support npu

csrc/ktransformers_ext/CMakeLists.txt

@@ -44,6 +44,10 @@ option(KTRANSFORMERS_USE_ROCM                "ktransformers: use ROCM"
 option(KTRANSFORMERS_USE_XPU                 "ktransformers: use XPU"                           OFF)
 option(KTRANSFORMERS_USE_NPU                 "ktransformers: use NPU"                           OFF)
 
+if(KTRANSFORMERS_USE_NPU)
+    add_definitions(-DKTRANSFORMERS_USE_NPU=1)
+endif()
+
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
@@ -90,6 +94,9 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
         endif ()
         set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
     else()
+        if(KTRANSFORMERS_USE_NPU)
+            list(APPEND ARCH_FLAGS -march=armv8.2-a+fp16+fp16fml+dotprod -lnuma)
+        endif()
         check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
         if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
             list(APPEND ARCH_FLAGS -mfp16-format=ieee)
@@ -117,37 +124,38 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
         (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
          CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$"))
     message(STATUS "x86 detected")
-    set(HOST_IS_X86 TRUE)
-    set(HAS_AVX512 TRUE)
-    set(__HAS_AMX__ TRUE)
-    add_compile_definitions(__x86_64__)
-    # check AVX512
-    execute_process(
-        COMMAND lscpu
-        OUTPUT_VARIABLE LSCPU_OUTPUT
-        OUTPUT_STRIP_TRAILING_WHITESPACE
-    )
-    # message(STATUS "LSCPU_OUTPUT: ${LSCPU_OUTPUT}")
+    if(NOT KTRANSFORMERS_USE_NPU)
+        set(HOST_IS_X86 TRUE)
+        set(HAS_AVX512 TRUE)
+        set(__HAS_AMX__ TRUE)
+        add_compile_definitions(__x86_64__)
+        # check AVX512
+        execute_process(
+            COMMAND lscpu
+            OUTPUT_VARIABLE LSCPU_OUTPUT
+            OUTPUT_STRIP_TRAILING_WHITESPACE
+        )
+        # message(STATUS "LSCPU_OUTPUT: ${LSCPU_OUTPUT}")
 
-    string(FIND "${LSCPU_OUTPUT}" "avx512" COMPILER_SUPPORTS_AVX512F)
-    
-    if (COMPILER_SUPPORTS_AVX512F GREATER -1)
-        message(STATUS "Compiler and CPU support AVX512F (tested by compiling a program)")
-        add_compile_definitions(__HAS_AVX512F__)
-    else()
-        message(STATUS "Compiler and/or CPU do NOT support AVX512F")
-        set(HAS_AVX512 False)
-    endif()
+        string(FIND "${LSCPU_OUTPUT}" "avx512" COMPILER_SUPPORTS_AVX512F)
+        
+        if (COMPILER_SUPPORTS_AVX512F GREATER -1)
+            message(STATUS "Compiler and CPU support AVX512F (tested by compiling a program)")
+            add_compile_definitions(__HAS_AVX512F__)
+        else()
+            message(STATUS "Compiler and/or CPU do NOT support AVX512F")
+            set(HAS_AVX512 False)
+        endif()
 
-    # check AMX
-    string(FIND "${LSCPU_OUTPUT}" "amx" COMPILER_SUPPORTS_AMX)
-    
-    if(COMPILER_SUPPORTS_AMX GREATER -1)
-        message(STATUS "Compiler supports AMX")
-        add_compile_definitions(__HAS_AMX__)
-    else()
-        message(STATUS "Compiler does NOT support AMX")
-    endif()
+        # check AMX
+        string(FIND "${LSCPU_OUTPUT}" "amx" COMPILER_SUPPORTS_AMX)
+        
+        if(COMPILER_SUPPORTS_AMX GREATER -1)
+            message(STATUS "Compiler supports AMX")
+            add_compile_definitions(__HAS_AMX__)
+        else()
+            message(STATUS "Compiler does NOT support AMX")
+        endif()
     if (MSVC)
         # instruction set detection for MSVC only
         if (LLAMA_NATIVE)
@@ -281,6 +289,8 @@ if (WIN32)
     include_directories("$ENV{CUDA_PATH}/include")
     add_compile_definitions(KTRANSFORMERS_USE_CUDA=1)
 elseif (UNIX)
+
+
     if (KTRANSFORMERS_USE_ROCM)
         find_package(HIP REQUIRED)
         if(HIP_FOUND)

ktransformers/local_chat_npu.py

@@ -0,0 +1,257 @@
+"""
+Description  :  
+Author       : Boxin Zhang, Azure-Tang
+Version      : 0.1.0
+Copyright (c) 2024 by KVCache.AI, All Rights Reserved. 
+"""
+
+import os
+import platform
+import sys
+
+project_dir = os.path.dirname(os.path.dirname(__file__))
+sys.path.insert(0, project_dir)
+import torch
+import torch_npu
+from torch_npu.contrib import transfer_to_npu
+import torch.distributed as dist
+
+import logging
+from transformers import (
+    AutoTokenizer,
+    AutoConfig,
+    AutoModelForCausalLM,
+    GenerationConfig,
+    TextStreamer,
+)
+import json
+import fire
+from ktransformers.optimize.optimize import optimize_and_load_gguf
+from ktransformers.models.modeling_deepseek import DeepseekV2ForCausalLM
+from ktransformers.models.modeling_qwen2_moe import Qwen2MoeForCausalLM
+from ktransformers.models.modeling_deepseek_v3 import DeepseekV3ForCausalLM
+from ktransformers.models.modeling_llama import LlamaForCausalLM
+from ktransformers.models.modeling_mixtral import MixtralForCausalLM
+from ktransformers.util.utils import prefill_and_generate, get_compute_capability
+from ktransformers.util.ascend.ascend_utils import get_absort_weight, setup_model_parallel, get_tensor_parallel_group
+from ktransformers.util import utils
+from ktransformers.models.custom_cache import StaticCache
+from ktransformers.server.config.config import Config
+from ktransformers.operators.flashinfer_wrapper import flashinfer_enabled
+from ktransformers.util.vendors import device_manager, get_device, to_device, GPUVendor
+
+custom_models = {
+    "DeepseekV2ForCausalLM": DeepseekV2ForCausalLM,
+    "DeepseekV3ForCausalLM": DeepseekV3ForCausalLM,
+    "Qwen2MoeForCausalLM": Qwen2MoeForCausalLM,
+    "LlamaForCausalLM": LlamaForCausalLM,
+    "MixtralForCausalLM": MixtralForCausalLM,
+}
+torch.npu.config.allow_internal_format = True
+
+ktransformer_rules_dir = (
+    os.path.dirname(os.path.abspath(__file__)) + "/optimize/optimize_rules/"
+)
+default_optimize_rules = {
+    "DeepseekV3ForCausalLM": ktransformer_rules_dir + "npu/DeepSeek-V3-Chat.yaml",
+}
+torch.npu.set_compile_mode(jit_compile=False)
+
+
+import sys, signal, faulthandler
+faulthandler.register(signal.SIGUSR1, file=sys.stderr, all_threads=True, chain=False)
+
+
+def local_chat(
+    model_path: str | None = None,
+    optimize_config_path: str = None,
+    gguf_path: str | None = None,
+    max_new_tokens: int = 1000,
+    cpu_infer: int = Config().cpu_infer,
+    use_cuda_graph: bool = False,
+    prompt_file : str | None = None,
+    mode: str = "normal",
+    force_think: bool = False,
+    chunk_size: int = utils._MAX_CHUNK_SIZE,
+    q4_gguf_path: str | None = None,
+    tp: int = 1,
+):
+    utils.USE_NPU_GRAPH = use_cuda_graph
+    torch.npu.config.allow_internal_format = False
+    torch.set_grad_enabled(False)
+    Config().cpu_infer = cpu_infer
+
+    local_rank, world_size = setup_model_parallel(tp=tp)
+    if utils.CUR_DEVICE is None:
+        utils.CUR_DEVICE = f"npu:{torch.npu.current_device()}"
+
+    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
+    if use_cuda_graph:
+        from ktransformers.util import npu_graph_runner
+        npu_graph_runner.LAYER_ID = config.num_hidden_layers
+    if mode == 'long_context':
+        assert config.architectures[0] == "LlamaForCausalLM", "only LlamaForCausalLM support long_context mode"
+        torch.set_default_dtype(torch.float16)
+    else:
+        torch.set_default_dtype(config.torch_dtype)
+
+    with torch.device("meta"):
+        if config.architectures[0] in custom_models:
+            print("using custom modeling_xxx.py.")
+            if (
+                "Qwen2Moe" in config.architectures[0]
+            ):  # Qwen2Moe must use flash_attention_2 to avoid overflow.
+                config._attn_implementation = "flash_attention_2"
+            if "Llama" in config.architectures[0]:
+                config._attn_implementation = "eager"
+            if "Mixtral" in config.architectures[0]:
+                config._attn_implementation = "flash_attention_2"
+
+            model = custom_models[config.architectures[0]](config)
+        else:
+            model = AutoModelForCausalLM.from_config(
+                config, trust_remote_code=True, attn_implementation="flash_attention_2"
+            )
+
+    if optimize_config_path is None:
+        if config.architectures[0] in default_optimize_rules:
+            print("using default_optimize_rule for", config.architectures[0]) if local_rank == 0 else None
+            optimize_config_path = default_optimize_rules[config.architectures[0]]
+            print(f'{optimize_config_path=}') if local_rank == 0 else None
+        else:
+            optimize_config_path = input(
+                "please input the path of your rule file(yaml file containing optimize rules):"
+            )
+
+    if gguf_path is None:
+        gguf_path = input(
+            "please input the path of your gguf file(gguf file in the dir containing input gguf file must all belong to current model):"
+        )
+    optimize_and_load_gguf(model, optimize_config_path, gguf_path, config, q4_gguf_path=q4_gguf_path)
+    get_absort_weight(model, config)
+
+    try:
+        model.generation_config = GenerationConfig.from_pretrained(model_path)
+    except Exception as e:
+        print(f"generation config can't auto create, make default. Message: {e}")
+        gen_config = GenerationConfig(
+            temperature=0.6,
+            top_p=0.95,
+            do_sample=True
+        )
+        model.generation_config = gen_config
+    # model.generation_config = GenerationConfig.from_pretrained(model_path)
+    if model.generation_config.pad_token_id is None:
+        model.generation_config.pad_token_id = model.generation_config.eos_token_id
+    model.eval()
+    logging.basicConfig(level=logging.INFO)
+
+    system = platform.system()
+    if system == "Windows":
+        os.system("cls") if local_rank == 0 else None
+    else:
+        os.system("clear") if local_rank == 0 else None
+
+    print(f"{model=}") if local_rank == 0 else None
+
+    batch_size, seq_length = 1, 1024
+    device_map = model.gguf_loader.tensor_device_map
+    static_cache = StaticCache(
+        config = model.config, max_batch_size = batch_size, max_cache_len = seq_length + max_new_tokens, device = device_map,
+        dtype = model.dtype
+    )
+    chunk_size = int(chunk_size)
+    new_chunk_size = min(max(chunk_size, 512), utils._MAX_CHUNK_SIZE)
+    if new_chunk_size != chunk_size:
+        chunk_size = new_chunk_size
+        print(f'[WARN] Chunk size reset to legal value between [512, {utils._MAX_CHUNK_SIZE}] which is {chunk_size}.')
+
+    torch.distributed.barrier()
+    while True:
+        if local_rank == 0:
+            try:
+                content = input("Chat: ").strip()
+            except KeyboardInterrupt:
+                dist.barrier()
+                print('Exit all ranks with KeyboardInterrupt!')
+                sys.exit(0)
+            if content.startswith('"""'):  # prefix """
+                # multi lines input
+                content = content[3:] + "\n"
+                while True:
+                    line = input("")
+                    if line.endswith('"""'):
+                        # end multi lines input
+                        line = line[:-3]  # suffix """
+                        if line:
+                            content += line + "\n"
+                        break
+                    else:
+                        content += line + "\n"
+
+            if content == "":
+                if prompt_file != None:
+                    content = open(prompt_file, "r").read()
+                else:
+                    continue
+            elif os.path.isfile(content):
+                f = open(content, "r")
+                content = f.readlines()
+                f.close()
+            else:
+                content = [f"{len(content)},{max_new_tokens},{content}"]
+        else:
+            content = [""]
+
+        for line in content:
+            content_tensor = torch.tensor(bytearray(line.encode()), dtype=torch.uint8).to(device=utils.CUR_DEVICE)
+            if world_size > 1:
+                content_size = torch.tensor(len(content_tensor), dtype=torch.int64).to(device=utils.CUR_DEVICE)
+                all_content_sizes = [torch.zeros((1,), dtype=torch.int64).to(device=utils.CUR_DEVICE) for _ in range(world_size)]
+                dist.barrier()
+                dist.all_gather(all_content_sizes, content_size)
+                max_content_size = max([size.item() for size in all_content_sizes])
+
+                padded_content_tensor = torch.zeros((max_content_size,), dtype=torch.uint8).to(device=utils.CUR_DEVICE)
+                padded_content_tensor[:len(content_tensor)] = content_tensor
+
+                all_content_tensors = [torch.zeros((max_content_size,), dtype=torch.uint8).to(device=utils.CUR_DEVICE) for _ in range(world_size)]
+                dist.barrier()
+                dist.all_gather(all_content_tensors, padded_content_tensor)
+                content_tensor = all_content_tensors[0][:all_content_sizes[0].item()]
+                line = bytes(content_tensor.cpu().numpy()).decode()
+
+            parts = line.split(",")
+            input_tokens = int(parts[0])
+            max_new_tokens = int(parts[1])
+            line = line[line.index(",", line.index(",") + 1) + 1:]
+            
+            messages = [{"role": "user", "content": line}]
+            input_tensor = tokenizer.apply_chat_template(
+                messages, add_generation_prompt=True, return_tensors="pt"
+            )
+            if force_think:
+                token_thinks = torch.tensor([tokenizer.encode("<think>\\n",add_special_tokens=False)],device=input_tensor.device)
+                input_tensor = torch.cat(
+                    [input_tensor, token_thinks], dim=1
+                )
+            if mode == 'long_context':
+                assert Config().long_context_config['max_seq_len'] > input_tensor.shape[1] + max_new_tokens, \
+                "please change max_seq_len in  ~/.ktransformers/config.yaml"
+
+            if system != "Windows" and (config.architectures[0] == "DeepseekV2ForCausalLM" or config.architectures[0] == "DeepseekV3ForCausalLM") and flashinfer_enabled and get_compute_capability() >= 8 and device_manager.gpu_vendor == GPUVendor.NVIDIA:
+                generated = prefill_and_generate(
+                    model, tokenizer, input_tensor.cuda(), max_new_tokens, use_cuda_graph, mode = mode, force_think = force_think, chunk_size = chunk_size,
+                    use_flashinfer_mla = True, num_heads = config.num_attention_heads, head_dim_ckv = config.kv_lora_rank, head_dim_kpe = config.qk_rope_head_dim, q_head_dim = config.qk_rope_head_dim + config.qk_nope_head_dim,
+                    static_cache=static_cache
+                )
+            else:
+                generated = prefill_and_generate(
+                    model, tokenizer, input_tensor.to(device=utils.CUR_DEVICE), max_new_tokens, use_cuda_graph, mode = mode, force_think = force_think, chunk_size = chunk_size,
+                    static_cache=static_cache
+                )
+
+
+if __name__ == "__main__":
+    fire.Fire(local_chat)

ktransformers/models/custom_cache.py

@@ -16,6 +16,16 @@ try:
     from ktransformers.server.balance_serve.settings import sched_ext
 except:
     print("no balance_serve")
+
+
+try:
+    import torch_npu
+    from ktransformers.util import utils
+
+    use_torch_npu = torch_npu.npu.is_available()
+except:
+    use_torch_npu = False
+
 class StaticCache(transformers.StaticCache):
     """
     Static Cache class to be used with `torch.compile(model)`.
@@ -37,6 +47,10 @@ class StaticCache(transformers.StaticCache):
     def __init__(self, config: PretrainedConfig, max_batch_size: int, max_cache_len: int, device: torch.device| dict, dtype=None) -> None:
         Cache.__init__(self)
         self.max_batch_size = max_batch_size
+        
+        if use_torch_npu:
+            self.position = [0]
+
         self.max_cache_len = config.max_position_embeddings if max_cache_len is None else max_cache_len
         # Some model define a custom `head_dim` != config.hidden_size // config.num_attention_heads
         if config.architectures[0] == "DeepseekV3ForCausalLM":
@@ -56,8 +70,18 @@ class StaticCache(transformers.StaticCache):
         cache_shape = (max_batch_size, self.num_key_value_heads, self.max_cache_len, self.head_dim)
         if config.architectures[0] == "DeepseekV2ForCausalLM" or config.architectures[0] == "DeepseekV3ForCausalLM":
             # TODO: for deepseek, cache_shape is different whether using Absorbed MLA, check it automatically
-            self.page_size = 64
-            self.max_pages = (self.max_cache_len + self.page_size - 1) // self.page_size
+
+            if use_torch_npu:
+                self.page_size = 128
+                self.page_size_tensor = torch.tensor(
+                    self.page_size,
+                    dtype=torch.int32,
+                ).npu()
+                self.max_pages_per_batch = (self.max_cache_len + self.page_size - 1) // self.page_size
+                self.max_pages = (self.max_cache_len + self.page_size - 1) // self.page_size * self.max_batch_size
+            else:
+                self.page_size = 64
+                self.max_pages = (self.max_cache_len + self.page_size - 1) // self.page_size
             latent_shape = (self.max_pages, self.page_size, 1, config.kv_lora_rank + config.qk_rope_head_dim)
             self.kv_lora_rank = config.kv_lora_rank
             self.qk_rope_head_dim = config.qk_rope_head_dim
@@ -71,9 +95,14 @@ class StaticCache(transformers.StaticCache):
                     target_device = device
                 
                 if target_device not in self.page_table_map:
-                    page_table = torch.zeros((max_batch_size, self.max_pages), dtype=torch.int32, device=target_device)
-                    for seq_id in range(max_batch_size):
-                        page_table[seq_id, :] = torch.arange(seq_id * self.max_pages, seq_id * self.max_pages + self.max_pages, dtype=torch.int32, device=target_device)
+                    if use_torch_npu:
+                        page_table = torch.zeros((max_batch_size, self.max_pages_per_batch), dtype=torch.int32, device=target_device)
+                        for seq_id in range(max_batch_size):
+                            page_table[seq_id, :] = torch.arange(seq_id * self.max_pages_per_batch, seq_id * self.max_pages_per_batch + self.max_pages_per_batch, dtype=torch.int32, device=target_device)
+                    else:
+                        page_table = torch.zeros((max_batch_size, self.max_pages), dtype=torch.int32, device=target_device)
+                        for seq_id in range(max_batch_size):
+                            page_table[seq_id, :] = torch.arange(seq_id * self.max_pages, seq_id * self.max_pages + self.max_pages, dtype=torch.int32, device=target_device)
                     self.page_table_map[target_device] = page_table
                     
                 self.page_table_list.append(self.page_table_map[target_device])
@@ -140,11 +169,24 @@ class StaticCache(transformers.StaticCache):
         self.past_tokens[layer_idx] += cache_position.size(0)
         #print(cache_position)
         if self.is_MLA:
-            page_idx = cache_position // self.page_size
-            page_offset = cache_position % self.page_size
-            # key shape (self.max_pages, self.page_size, 1, config.kv_lora_rank + config.qk_rope_head_dim)
-            k_out[page_idx, page_offset, :, :self.kv_lora_rank] = key_states
-            k_out[page_idx, page_offset, :, self.kv_lora_rank:] = value_states
+            if use_torch_npu:
+                page_idx = cache_position // self.page_size_tensor
+                page_offset = cache_position % self.page_size_tensor
+
+                page_idx = page_idx.unsqueeze(0).expand(self.max_batch_size, -1)
+                page_offset = page_offset.unsqueeze(0).expand(self.max_batch_size, -1)
+
+                page_idx_offset = torch.arange(self.max_batch_size, device=page_idx.device) * self.max_pages_per_batch
+                page_idx = page_idx + page_idx_offset.unsqueeze(1)
+
+                combined = torch.cat([key_states, value_states], dim=-1)
+                combined = combined.contiguous()
+            else:
+                page_idx = cache_position // self.page_size
+                page_offset = cache_position % self.page_size
+                # key shape (self.max_pages, self.page_size, 1, config.kv_lora_rank + config.qk_rope_head_dim)
+                k_out[page_idx, page_offset, :, :self.kv_lora_rank] = key_states
+                k_out[page_idx, page_offset, :, self.kv_lora_rank:] = value_states
             return k_out, self.page_table_list[layer_idx]
         else:
             k_out[:, :, cache_position] = key_states
@@ -178,6 +220,9 @@ class StaticCache(transformers.StaticCache):
             if self.value_cache[layer_idx] is not None:
                 self.value_cache[layer_idx].zero_()
             self.past_tokens[layer_idx] = 0
+        
+        if use_torch_npu:
+            self.position = [0]
 
     def remove_suffix(self, start_pos):
         for layer_idx in range(len(self.key_cache)):

ktransformers/operators/attention.py

@@ -27,8 +27,12 @@ try:
     from flash_attn import flash_attn_func
 except:
     pass
-from ktransformers.operators.triton_attention import decode_attention_fwd_grouped 
-from ktransformers.operators.triton_attention_prefill import context_attention_fwd
+try:
+    from ktransformers.operators.triton_attention import decode_attention_fwd_grouped 
+    from ktransformers.operators.triton_attention_prefill import context_attention_fwd
+except:
+    Warning("triton not found, if you are using npu, ignore this.")
+
 import os
 from ktransformers.operators.flashinfer_wrapper import flashinfer_enabled
 if flashinfer_enabled:

ktransformers/operators/flashinfer_batch_prefill_wrapper.py

@@ -1,5 +1,8 @@
 import torch
-import flashinfer
+try:
+    import flashinfer
+except:
+	Warning("flashinfer not found, if you are using npu, ignore this.")
 import gc
 try:
     from flash_attn import flash_attn_with_kvcache

ktransformers/operators/flashinfer_wrapper.py

@@ -5,7 +5,11 @@ Version      : 0.2.3
 '''
 import torch
 import os
-from ktransformers.operators.triton_attention import decode_attention_fwd_grouped
+
+try:
+    from ktransformers.operators.triton_attention import decode_attention_fwd_grouped
+except:
+	Warning("triton not found, if you are using npu, ignore this.")
 
 flashinfer_enabled = False
 

ktransformers/operators/linear.py

@@ -14,7 +14,15 @@ Copyright (c) 2024 by KVCache.AI, All Rights Reserved.
 import ctypes
 import torch
 from torch import Tensor, nn
-if not torch.xpu.is_available():
+
+try:
+    import torch_npu
+
+    use_torch_npu = torch_npu.npu.is_available()
+except:
+    use_torch_npu = False
+
+if not torch.xpu.is_available() and not use_torch_npu:
     import KTransformersOps
     import vLLMMarlin
 from ktransformers.util.custom_loader import GGUFLoader, SafeTensorLoader

ktransformers/optimize/optimize.py

@@ -16,6 +16,7 @@ from ktransformers.util.custom_loader import GGUFLoader, ModelLoaderFactory
 from ktransformers.util.utils import set_module, load_weights
 import itertools
 import copy
+from ktransformers.util import utils
 
 def inject(module, local_optimization_dict, model_config:AutoConfig ,gguf_loader:GGUFLoader, prefix=''):
     for name, child in module._modules.items():
@@ -114,7 +115,7 @@ def translate_model_config(model_config: PretrainedConfig):
     return model_config
 
 
-def optimize_and_load_gguf(module: nn.Module, rule_file: str, gguf_path: str, model_config: PretrainedConfig, default_device: str = "cuda:0"):
+def optimize_and_load_gguf(module: nn.Module, rule_file: str, gguf_path: str, model_config: PretrainedConfig, default_device: str = "cuda:0", q4_gguf_path=""):
     with open(rule_file, 'r', encoding='utf-8') as f:
         rule_list = yaml.load(f.read(), Loader=yaml.FullLoader)
     
@@ -123,15 +124,29 @@ def optimize_and_load_gguf(module: nn.Module, rule_file: str, gguf_path: str, mo
     
     model_config = translate_model_config(model_config)
 
-    weights_loader = ModelLoaderFactory.create_loader(gguf_path)
-    with torch.device("meta"):
-        inject(module, optimize_config, model_config, weights_loader)
-    # pre load lm_head because its big inter result
-    load_weights(module.lm_head, weights_loader, "lm_head.", device=default_device)
-    load_weights(module, weights_loader, device=default_device)
-    module.gguf_loader = weights_loader
+    if q4_gguf_path:
+        q4_gguf_loader = GGUFLoader(q4_gguf_path)
+        utils.Q4_GGUF_LODER = q4_gguf_loader
+        gguf_loader = GGUFLoader(gguf_path, getattr(model_config, "quantize", None))
+        with torch.device("meta"):
+            inject(module, optimize_config, model_config, gguf_loader)
+        # pre load lm_head because its big inter result
+        load_weights(module.lm_head, gguf_loader, "lm_head.")
+        load_weights(module, gguf_loader)
+        module.gguf_loader = gguf_loader
+
+    else:
+        weights_loader = ModelLoaderFactory.create_loader(gguf_path)
+        with torch.device("meta"):
+            inject(module, optimize_config, model_config, weights_loader)
+        # pre load lm_head because its big inter result
+        load_weights(module.lm_head, weights_loader, "lm_head.", device=default_device)
+        load_weights(module, weights_loader, device=default_device)
+        module.gguf_loader = weights_loader
     del_meta(module)
     if torch.cuda.is_available():
         torch.cuda.empty_cache()
     elif torch.xpu.is_available():
         torch.xpu.empty_cache()
+    else:
+        torch.cuda.empty_cache()

ktransformers/optimize/optimize_rules/npu/DeepSeek-V3-Chat-npu.yaml

@@ -0,0 +1,76 @@
+- match:
+    class: ktransformers.models.modeling_deepseek_v3.DeepseekV3RotaryEmbedding
+  replace:
+    class: ktransformers.operators.RoPE.YarnRotaryEmbeddingV3
+    kwargs:
+      generate_device: "npu"
+      prefill_device: "npu"
+
+- match:
+    name: "^lm_head$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "npu"
+      prefill_device: "npu"
+      generate_op: "KLinearTorch"
+      prefill_op: "KLinearTorch"
+
+- match:
+    name: "^model\\.layers\\.(?!.*self_attn\\.kv_b_proj).*$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "npu"
+      prefill_device: "npu"
+      generate_op: "KLinearTorch"
+      prefill_op: "KLinearTorch"
+- match:
+    name: "^model\\.layers\\..*\\.mlp$"
+    class: ktransformers.models.modeling_deepseek_v3.DeepseekV3MoE
+  replace:
+    class: ktransformers.operators.experts.KDeepseekV3MoE     # mlp module with custom forward function
+    kwargs:
+      generate_device: "npu"
+      prefill_device: "npu"
+- match:
+    class: ktransformers.models.modeling_deepseek_v3.MoEGate
+  replace:
+    class: ktransformers.operators.gate.KMoEGate
+    kwargs:
+      generate_device: "npu:0"
+      prefill_device: "npu:0"
+- match:
+    name: "^model\\.layers\\..*\\.mlp\\.experts$"
+  replace:
+    class: ktransformers.operators.experts.KTransformersExperts     # custom MoE Kernel with expert paralleism
+    kwargs:
+      prefill_device: "npu"
+      prefill_op: "KExpertsTorch"
+      generate_device: "cpu"
+      generate_op: "KExpertsCPU"
+      out_device: "npu"
+  recursive: False # don't recursively inject submodules of this module
+- match:
+    name: "^model\\.layers\\..*\\.self_attn$"
+  replace:
+    class: ktransformers.operators.attention.KDeepseekV2Attention # optimized MLA implementation
+    kwargs:
+      generate_device: "npu"
+      prefill_device: "npu"
+      absorb_for_prefill: False # change this to True to enable long context(prefill may slower).
+- match:
+    name: "^model$"
+  replace:
+    class: "ktransformers.operators.models.KDeepseekV2Model"
+    kwargs:
+      per_layer_prefill_intput_threshold: 0 # 0 is close layer wise prefill
+- match:
+    name: "^model.embed_tokens"
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cpu"
+      prefill_device: "cpu"

ktransformers/server/args.py

@@ -22,6 +22,10 @@ class ArgumentParser:
             "--device", type=str, default=self.cfg.model_device, help="Warning: Abandoning this parameter"
         )
         parser.add_argument("--architectures", type=str, default=self.cfg.model_name)
+
+        parser.add_argument("--tp", type=int, default=1)
+        parser.add_argument("--q4_gguf_path", type=str, default=None)
+
         parser.add_argument("--gguf_path", type=str, default=self.cfg.gguf_path)
         parser.add_argument("--optimize_config_path", default=None, type=str, required=False)
         parser.add_argument("--cpu_infer", type=int, default=self.cfg.cpu_infer)

ktransformers/server/backend/args.py

@@ -8,6 +8,7 @@ class ConfigArgs(BaseModel):
     model_dir: Optional[str] = Field(..., description="Path to model directory")
     optimize_config_path: Optional[str] = Field(None, description="Path of your optimize config yml file")
     gguf_path: Optional[str] = Field(None, description="Path of your gguf file")
+    tp: int = Field(None, description="tp size")
 
     class Config:
         protected_namespaces = ()

ktransformers/server/backend/interfaces/ktransformers.py

@@ -1,4 +1,19 @@
 import torch
+from torch import nn
+
+try:
+    import torch_npu
+    from ktransformers.util.ascend.ascend_utils import get_absort_weight, setup_model_parallel
+    from ktransformers.util.utils import get_device, get_all_used_cuda_device
+    from ktransformers.util import utils
+
+    use_torch_npu = torch_npu.npu.is_available()
+except:
+    use_torch_npu = False
+
+import os
+
+
 from typing import Optional, List
 import asyncio
 from transformers import AutoTokenizer, AutoConfig, GenerationConfig
@@ -19,6 +34,9 @@ from typing import Optional
 from ktransformers.operators.flashinfer_wrapper import flashinfer_enabled, MLAWrapperSingleton
 from ktransformers.server.schemas.endpoints.chat import RawUsage
 
+
+
+
 warm_uped = False
 
 class KTransformersThreadContext(TransformersThreadContext):
@@ -26,8 +44,15 @@ class KTransformersThreadContext(TransformersThreadContext):
 
 
 class KTransformersInterface(TransformersInterface):
-    def __init__(self, args: ConfigArgs = default_args):
-        self.args = args
+    def __init__(self, args: ConfigArgs = default_args, input_args=None):
+        if use_torch_npu:
+            self.args = input_args
+            self.local_rank, self.world_size = setup_model_parallel(tp=self.args.tp)
+            if utils.CUR_DEVICE is None:
+                utils.CUR_DEVICE = f"npu:{torch.npu.current_device()}"
+            self.args.device = utils.CUR_DEVICE            
+        else:
+            self.args = args
         torch.set_grad_enabled(False)
         self.tokenizer = AutoTokenizer.from_pretrained(args.model_dir, device=args.device, trust_remote_code=args.trust_remote_code)
         config = AutoConfig.from_pretrained(args.model_dir, trust_remote_code=args.trust_remote_code)
@@ -47,7 +72,10 @@ class KTransformersInterface(TransformersInterface):
 
         with torch.device("meta"):
             self.model = custom_models[config.architectures[0]](config)
-        if default_args.optimize_config_path is None:
+
+        if use_torch_npu and input_args.optimize_config_path is not None:
+            optimize_config_path = input_args.optimize_config_path
+        elif default_args.optimize_config_path is None:
             optimize_config_path = default_optimize_rules[config.architectures[0]]
         else:
             optimize_config_path = args.optimize_config_path
@@ -60,7 +88,14 @@ class KTransformersInterface(TransformersInterface):
                 "please input the path of your gguf file(gguf file in the dir containing input gguf file must all"
                 " belong to current model):"
             )
-        optimize_and_load_gguf(self.model, optimize_config_path, gguf_path, config)
+
+        if use_torch_npu:
+            optimize_and_load_gguf(self.model, optimize_config_path, gguf_path, config, q4_gguf_path=input_args.q4_gguf_path)
+            #提前absorbed
+            get_absort_weight(self.model, config)
+            self.model.eval()
+        else:
+            optimize_and_load_gguf(self.model, optimize_config_path, gguf_path, config)
         self.model.generation_config = generation_config
         self.device_map = self.model.gguf_loader.tensor_device_map
         # logger.info(f"{args.model_name} loaded from {args.model_dir} to {self.device_map}")
@@ -77,9 +112,92 @@ class KTransformersInterface(TransformersInterface):
             self.model.generation_config.pad_token_id = self.model.generation_config.eos_token_id
         self.streamer = TextStreamer(self.tokenizer)
 
+        if use_torch_npu:
+            self.top_p = torch.tensor([[self.model.generation_config.top_p]], dtype=torch.float16, device=self.args.device)
+            self.top_k = torch.tensor([[self.model.generation_config.top_k]], dtype=torch.int32, device=self.args.device)
+            self.temperature = torch.tensor([[self.model.generation_config.temperature]], dtype=torch.float16, device=self.args.device)
+            self.next_token_fake = torch.tensor([[1]], dtype=torch.int32, device=self.args.device)
+            self.next_token_probs = torch.tensor([[1.0]], dtype=torch.float16, device=self.args.device)
+            self._infer_lock = asyncio.Lock()
+
+
         self._infer_lock = asyncio.Lock()
 
+    def decode_logits_to_token(self, logits: torch.Tensor):
+        if self.model.generation_config.do_sample:
+            logits = logits / self.temperature
+            torch.manual_seed(0)
+            probs = logits.view(1, self.model.config.vocab_size)
+            sm = nn.Softmax(dim=-1)
+            probs = sm(probs).half().npu()
+            next_token = self.next_token_fake
+            torch_npu._npu_topk_topp_sampling(probs, self.top_k, self.top_p, next_token, self.next_token_probs)
+            last = next_token.squeeze(-1)
+        else:
+            logits = self.logits_warper(self.inputs.view(1, -1), logits.view(1, -1))
+            probs = torch.nn.functional.softmax(logits, dim=-1)
+            _, last = torch.topk(probs, k=1, dim=-1)
+        last = last.item()
+        self.ever_generated_ids.add(last)
+        return last
+
+    def decode_one_tokens_npu(self):
+        global warm_uped
+
+        device_map = self.model.gguf_loader.tensor_device_map
+        torch_device = get_device("blk.0.self_attn", device_map)
+        torch_device = "cuda:0" if torch_device == "cuda" else torch_device
+        torch.cuda.set_device(torch_device)
+        if warm_uped and self.args.use_cuda_graph:
+            from ktransformers.util.npu_graph_runner import get_or_create_runner, check_runner
+            if check_runner(self.args.device):
+                npu_graph_runner = get_or_create_runner(self.args.device)
+                npu_graph_runner.init(self.args.batch_size, self.seq_length)
+                self.cuda_graph_runner = npu_graph_runner
+                utils._USE_NPU_GRAPH = True
+                self.cuda_graph_runner.capture(
+                    self.model,
+                    self.current_ids,
+                    self.active_cache_position.unsqueeze(0),
+                    self.active_cache_position,
+                    self.cache,
+                    main_device=self.args.device,
+                    return_dict=False,
+                    use_cache=True,
+                )
+
+            if hasattr(self, "cuda_graph_runner"):
+                inputs_embeds = self.model.model.embed_tokens(self.current_ids.to("cpu")).to(self.args.device)
+                logits = self.cuda_graph_runner(
+                    inputs_embeds, self.active_cache_position.unsqueeze(0), self.active_cache_position
+                )
+                self.cache.change_seq_length(1)
+                torch.cuda.synchronize()
+                logits = logits[0, -1, :]
+                return self.decode_logits_to_token(logits)
+        
+        if self.args.use_cuda_graph:
+            warm_uped = True
+            
+        if self.use_static_cache:
+            logits = self.model(
+                self.current_ids.to(torch_device),
+                cache_position=self.active_cache_position,
+                past_key_values=self.cache,
+                return_dict=False,
+                use_cache=True,
+            )[0]
+        else:
+            logits = self.model(self.current_ids, return_dict=False)[0]
+        self.cache.change_seq_length(1)
+        logits = logits[0, -1, :]
+
+        return self.decode_logits_to_token(logits)
+
     def decode_one_tokens(self):
+        if use_torch_npu:
+            return self.decode_one_tokens_npu()
+
         global warm_uped
 
         device_map = self.model.gguf_loader.tensor_device_map
@@ -127,9 +245,145 @@ class KTransformersInterface(TransformersInterface):
         return self.logits_to_token(logits)
 
 
+    @torch.no_grad
+    def prefill_npu(self, input_ids: torch.Tensor, is_new: bool, temperature: Optional[float] = None, top_p: Optional[float] = None, max_tokens: Optional[float] = None, max_completion_tokens: Optional[float] = None):
+        input_ids_length = input_ids.shape[-1]
+        if(input_ids_length >= self.args.cache_lens):
+            logger.warning(f"input_ids_length {input_ids_length} > cache_lens {self.args.cache_lens}")
+            self.seq_length = input_ids_length
+            return
+        logger.debug(f"input_ids: {input_ids.shape}")
+        device = self.device_map.get("blk.0.self_attn", {}).get("generate_device", "cuda:0")
+        device = "cuda:0" if device == "cuda" else device
+        device = self.args.device
+        if is_new:
+            self.ever_generated_ids.clear()
+            same_prefix = 0
+            flat_input_ids = input_ids.flatten()
+
+            if getattr(self, 'generated_ids', None) is None:
+                self.generated_ids = torch.zeros(
+                    self.args.batch_size,
+                    input_ids.shape[-1] + self.args.max_new_tokens + 1,
+                    dtype=torch.int,
+                    device=self.args.device,
+                )
+                self.seq_length = 1
+            
+            # flat_prev_ids = self.generated_ids.flatten()
+            # for i in range(min(self.seq_length, flat_input_ids.shape[0]) - 1):
+            #     if flat_input_ids[i] == flat_prev_ids[i]:
+            #         same_prefix += 1
+            #     else:
+            #         break
+
+            logger.debug(f"same prefix len: {same_prefix}")
+            self.cache.remove_suffix(same_prefix)
+            self.seq_length = same_prefix
+            self.cache.position[0] = same_prefix
+            self.generated_ids = self.generated_ids[..., :same_prefix]
+            input_ids = input_ids[..., same_prefix:]
+            input_ids_length = input_ids.shape[-1]
+
+        self.ever_generated_ids.clear()
+        self.profiler.set_counter("prefill", input_ids_length)
+        logger.debug(f"input_ids: {input_ids.shape}")
+        logger.debug(f"generate_ids: {self.generated_ids.shape}")
+        
+        former_seq_length = self.seq_length
+        self.seq_length += input_ids_length
+        expected_length = min(self.seq_length + self.args.max_new_tokens + 1, self.args.cache_lens)
+        delta_length = expected_length - self.generated_ids.shape[-1]
+        if delta_length > 0:
+            new_generate_ids = torch.zeros(
+                self.args.batch_size, delta_length, dtype=torch.int, device=self.args.device
+            )
+            self.generated_ids = torch.cat([self.generated_ids, new_generate_ids], dim=-1)
+        else:
+            logger.warning(f"seq_length bigger than cache_lens, killed")
+            exit(0)
+        
+        logger.debug(f"cache position: {former_seq_length} to {self.seq_length}")
+        cache_position = torch.arange(former_seq_length, self.seq_length, device=device)
+        self.cache.position[0] = self.seq_length + 1
+        self.generated_ids[:, cache_position] = input_ids.to(self.args.device).to(torch.int)
+
+        if not (type(self) is TransformersInterface):
+            input_ids = input_ids.to("cpu")
+        
+        def chunk_prefill(input_ids, cache_position):
+            inputs_embeds = self.model.model.embed_tokens(input_ids).to(device)
+            torch.cuda.set_device(device)
+            if flashinfer_enabled:
+                MLAWrapperSingleton.need_plan_all()
+            if self.use_static_cache:
+                logits = self.model(
+                    inputs_embeds=inputs_embeds,
+                    cache_position=cache_position,
+                    past_key_values=self.cache,
+                    return_dict=False,
+                    use_cache=True,
+                )[0]
+            else:
+                logits = self.model(inputs_embeds=inputs_embeds, return_dict=False)[0]
+
+            return logits
+
+        logits = None
+        def prefill_wrapper(prof=None):
+            nonlocal logits
+            chunk_start = 0
+            while chunk_start < input_ids_length:
+                chunk_end = min(chunk_start + self.args.chunk_size, input_ids_length)
+                if self.cache != None:
+                    self.cache.cur_idx = cache_position[chunk_start:chunk_end]
+                logits = chunk_prefill(input_ids[:, chunk_start:chunk_end], cache_position[chunk_start:chunk_end])
+                chunk_start += self.args.chunk_size
+                if prof is not None:
+                    prof.step()
+            if prof is not None:
+                prof.stop()
+            if logits is None:
+                raise ValueError('logits cannot be None')
+
+
+        global WARM_UP_SKIP_CNT
+        prof_prefill = os.environ["PROF_PREFILL"] if "PROF_PREFILL" in os.environ else "0"
+        if prof_prefill == "1":
+            experimental_config = torch_npu.profiler._ExperimentalConfig(
+                aic_metrics=torch_npu.profiler.AiCMetrics.PipeUtilization,
+                profiler_level=torch_npu.profiler.ProfilerLevel.Level1, l2_cache=False
+            )
+            with torch_npu.profiler.profile(
+                    activities=[
+                        torch_npu.profiler.ProfilerActivity.CPU,
+                        torch_npu.profiler.ProfilerActivity.NPU
+                    ],
+                    schedule=torch_npu.profiler.schedule(wait=0, warmup=0, active=8, repeat=1, skip_first=0),
+                    on_trace_ready=torch_npu.profiler.tensorboard_trace_handler("./prefill_prof_lm_head"),
+                    record_shapes=True,
+                    profile_memory=True,
+                    with_stack=False,
+                    with_flops=False,
+                    with_modules=False,
+                    experimental_config=experimental_config) as prof:
+                prefill_wrapper(prof)
+        else:
+            prefill_wrapper()
+            
+        if flashinfer_enabled:
+            MLAWrapperSingleton.reset_buffer()
+        self.prepare_logits_wrapper(input_ids, device, temperature, top_p)
+        next_token = self.logits_to_token(logits[0, -1, :])
+        yield self.append_new_tokens(next_token)
+
 
     @torch.no_grad
     def prefill(self, input_ids: torch.Tensor, is_new: bool, temperature: Optional[float] = None, top_p: Optional[float] = None, max_tokens: Optional[float] = None, max_completion_tokens: Optional[float] = None):
+
+        if use_torch_npu:
+            return self.prefill_npu(self, input_ids, is_new, temperature, top_p, max_tokens, max_completion_tokens)
+
         input_ids_length = input_ids.shape[-1]
         if max_tokens is not None:
             max_completion_tokens = max_tokens
@@ -144,6 +398,8 @@ class KTransformersInterface(TransformersInterface):
         logger.debug(f"input_ids: {input_ids.shape}")
         device = self.device_map.get("blk.0.self_attn", {}).get("generate_device", "cuda:0")
         device = "cuda:0" if device == "cuda" else device
+        if use_torch_npu:
+            device = self.args.device
 
         if is_new:
             self.ever_generated_ids.clear()
@@ -159,16 +415,19 @@ class KTransformersInterface(TransformersInterface):
                 )
                 self.seq_length = 1            
             
-            flat_prev_ids = self.generated_ids.flatten()
-            for i in range(min(self.seq_length, flat_input_ids.shape[0]) - 1):
-                if flat_input_ids[i] == flat_prev_ids[i]:
-                    same_prefix += 1
-                else:
-                    break
+            if not use_torch_npu:
+                flat_prev_ids = self.generated_ids.flatten()
+                for i in range(min(self.seq_length, flat_input_ids.shape[0]) - 1):
+                    if flat_input_ids[i] == flat_prev_ids[i]:
+                        same_prefix += 1
+                    else:
+                        break
             
             logger.debug(f"same prefix len: {same_prefix}")
             self.cache.remove_suffix(same_prefix)
             self.seq_length = same_prefix
+            if use_torch_npu:
+                self.cache.position[0] = same_prefix
             self.generated_ids = self.generated_ids[..., :same_prefix]
             input_ids = input_ids[..., same_prefix:]
             input_ids_length = input_ids.shape[-1]
@@ -193,6 +452,8 @@ class KTransformersInterface(TransformersInterface):
         
         logger.debug(f"cache position: {former_seq_length} to {self.seq_length}")
         cache_position = torch.arange(former_seq_length, self.seq_length, device=device)
+        if use_torch_npu:
+            self.cache.position[0] = self.seq_length + 1
         self.generated_ids[:, cache_position] = input_ids.to(self.args.device).to(torch.int)
 
         if not (type(self) is TransformersInterface):
@@ -248,4 +509,18 @@ class KTransformersInterface(TransformersInterface):
                 decode_time = self.profiler.get_timer_sec('decode'),
                 prefill_count = self.profiler.get_counter('prefill'),
                 decode_count = self.profiler.get_counter('decode'),
-            )
+            )
+
+    def sync_inference(self, local_messages, thread_id: str, temperature: Optional[float] = None, top_p: Optional[float] = None) -> str:
+        loop = asyncio.new_event_loop()
+        asyncio.set_event_loop(loop)
+
+        try:
+            async def run_async():
+                result = []
+                async for chunk in self.inference(local_messages, thread_id, temperature, top_p):
+                    pass
+                return ""
+            return loop.run_until_complete(run_async())
+        finally:
+            loop.close()

ktransformers/server/backend/interfaces/transformers.py

@@ -32,6 +32,20 @@ from ktransformers.server.config.log import logger
 from ..args import ConfigArgs, default_args
 from ktransformers.operators.flashinfer_wrapper import flashinfer_enabled, MLAWrapperSingleton
 
+
+
+try:
+    import torch_npu
+    from ktransformers.util import utils
+
+    use_torch_npu = torch_npu.npu.is_available()
+except:
+    use_torch_npu = False
+
+
+import torch.distributed as dist
+
+
 # This TextStreamer is a modified version from https://github.com/huggingface/transformers/blob/main/src/transformers/generation/streamers.py
 class TextStreamer:
 
@@ -191,11 +205,19 @@ class TransformersInterface(BackendInterfaceBase):
         #     input_ids = self.tokenizer.apply_chat_template(
         #         new_messages, return_tensors="pt", add_generation_prompt=True
         #     ).to(self.args.device)
-        input_str: str = self.tokenizer.apply_chat_template(new_messages,tokenize=False,add_generation_prompt=True)
-        # drop <think> token in chat template
-        if input_str.endswith('<think>\n'):
-            input_str = input_str[:-len('<think>\n')]
-        input_ids = self.tokenizer.encode(input_str, return_tensors="pt").to(self.args.device)
+
+        if not use_torch_npu:
+            input_str: str = self.tokenizer.apply_chat_template(new_messages,tokenize=False,add_generation_prompt=True)
+            # drop <think> token in chat template
+            if input_str.endswith('<think>\n'):
+                input_str = input_str[:-len('<think>\n')]
+            input_ids = self.tokenizer.encode(input_str, return_tensors="pt").to(self.args.device)
+        else:
+            logger.debug(f"new_messages: {new_messages}")
+            input_ids = self.tokenizer.apply_chat_template(
+                new_messages, add_generation_prompt=True, return_tensors="pt"
+            ) 
+
         if (self.last_request_id is not None) and self.last_request_id == thread_id:
             x = self.generated_ids[:,:self.seq_length]
             y = input_ids[:,:self.seq_length]
@@ -212,6 +234,8 @@ class TransformersInterface(BackendInterfaceBase):
     def append_new_tokens(self, new_tokens: int) -> Optional[str]:
         self.generated_ids[0, self.seq_length] = new_tokens
         self.seq_length += 1
+        if use_torch_npu:
+            self.cache.position[0] = self.seq_length
         return self.streamer.put(new_tokens)
 
     @staticmethod
@@ -273,14 +297,21 @@ class TransformersInterface(BackendInterfaceBase):
             top_p = self.model.generation_config.top_p
         if top_p == 0:
             top_p = 0.0001
-        generation_config, model_kwargs = self.model._prepare_generation_config(
-            None, max_length=self.args.max_new_tokens,
-            do_sample=True, 
-            top_k=self.args.top_k, 
-            top_p=top_p, 
-            temperature=temperature,
-            repetition_penalty=self.args.repetition_penalty # change this to modify generate config
-        )
+        
+        if use_torch_npu:
+            generation_config, model_kwargs = self.model._prepare_generation_config(
+                None, do_sample=True,
+                top_p=top_p, temperature=temperature
+            )
+        else:
+            generation_config, model_kwargs = self.model._prepare_generation_config(
+                None, max_length=self.args.max_new_tokens,
+                do_sample=True, 
+                top_k=self.args.top_k, 
+                top_p=top_p, 
+                temperature=temperature,
+                repetition_penalty=self.args.repetition_penalty # change this to modify generate config
+            )
         self.inputs = inputs
 
         self.logits_warper = self.tf_logits_warper(generation_config)
@@ -372,7 +403,10 @@ class TransformersInterface(BackendInterfaceBase):
         cache_position = torch.arange(former_seq_length, self.seq_length, device=self.args.device)
         self.generated_ids[:, cache_position] = input_ids.to(self.args.device).to(torch.int)
 
-        device = input_ids.device
+        if use_torch_npu:
+            device = self.args.device
+        else:
+            device = input_ids.device
         if not (type(self) is TransformersInterface):
             input_ids = input_ids.to("cpu")
         inputs_embeds = self.model.model.embed_tokens(input_ids).to(device)
@@ -420,7 +454,12 @@ class TransformersInterface(BackendInterfaceBase):
         else:   # for's else, if output get max new tokens
             yield self.streamer.end(), None
             yield "", "length"
-        
+    
+        if use_torch_npu and self.args.use_cuda_graph:
+            utils._USE_NPU_GRAPH = False
+            from ktransformers.util.npu_graph_runner import get_or_create_runner
+            npu_graph_runner = get_or_create_runner(self.args.device)
+            npu_graph_runner.destroy()
         
 
     def check_is_new(self, thread_id: str):
@@ -436,7 +475,87 @@ class TransformersInterface(BackendInterfaceBase):
                 self.last_request_id = thread_id
                 return True
 
+
+    async def inference_npu(self, local_messages, thread_id: str, temperature: Optional[float] = None, top_p: Optional[float] = None):
+        self.streamer.reset()
+        self.profiler.create_and_start_timer("tokenize")
+        rank = torch.distributed.get_rank()
+        tp_size = utils.get_tensor_parallel_size()
+        world_size = torch.distributed.get_world_size()
+        if isinstance(local_messages, List):
+            input_ids = self.format_and_tokenize_input_ids(thread_id, local_messages)
+        elif isinstance(local_messages, str):
+            #local_messages = local_messages[0]['content']
+            input_ids = self.tokenize_prompt(local_messages)
+            #input_ids = torch.tensor([[6366]], device=input_ids.device)
+        else:
+            raise ValueError("local_messages should be List or str")
+
+        if tp_size == world_size and tp_size > 1:
+            torch.distributed.barrier()
+            input_size = torch.tensor([input_ids.size(1)], dtype=torch.int64, device=self.args.device)
+            all_input_sizes = [torch.zeros_like(input_size) for _ in range(world_size)]
+            dist.all_gather(all_input_sizes, input_size)
+
+            max_input_size = max([size.item() for size in all_input_sizes])
+            padded_input_ids = torch.zeros(1, max_input_size, dtype=input_ids.dtype, device=self.args.device)
+            padded_input_ids[0, :input_ids.size(1)] = input_ids[0]
+
+            all_padded_inputs = [torch.zeros_like(padded_input_ids) for _ in range(world_size)]
+            dist.all_gather(all_padded_inputs, padded_input_ids)
+
+            original_size = all_input_sizes[0].item()
+            input_ids = all_padded_inputs[0][:, :original_size]
+        
+        if Config().user_force_think:
+            token_thinks = torch.tensor([self.tokenizer.encode("<think>\n",add_special_tokens=False)],device=input_ids.device)
+            if not torch.equal(input_ids[0, -token_thinks.shape[-1]:], token_thinks[-1]):
+                input_ids = torch.cat(
+                    [input_ids, token_thinks], dim=1
+                )
+
+        self.profiler.pause_timer("tokenize")
+
+        self.profiler.create_and_start_timer("prefill")
+
+        if Config().user_force_think:
+            think = '<think>\n'
+            if tp_size == world_size and rank != 0:
+                pass
+            else:
+                print(think, end="",flush=True)
+            yield think, None
+        
+        for t in self.prefill(input_ids, self.check_is_new(thread_id), temperature, top_p):
+            # output think token after prefill done
+            if t is not None:
+                print(t, end="",flush=True)
+                yield t, None
+        self.profiler.pause_timer("prefill")
+
+        self.profiler.create_and_start_timer("decode")
+        for t, finish_reason in self.generate():
+            if t is not None:
+                if tp_size == world_size and rank != 0:
+                    pass
+                else:
+                    print(t, end="",flush=True)
+                yield t, finish_reason
+
+        if tp_size == world_size and rank != 0:
+            pass
+        else:
+            self.profiler.pause_timer("decode")
+            self.report_last_time_performance()
+
     async def inference(self, local_messages, thread_id: str, temperature: Optional[float] = None, top_p: Optional[float] = None, max_tokens: Optional[float] = None, max_completion_tokens: Optional[float] = None):
+
+        if use_torch_npu:
+            async for tok in self.inference_npu(local_messages, thread_id, temperature, top_p):
+                yield tok
+            return
+
+
         self.streamer.reset()
         self.profiler.create_and_start_timer("tokenize")
         if isinstance(local_messages, List):

ktransformers/server/main.py

@@ -9,7 +9,7 @@ project_dir = os.path.dirname(os.path.dirname(os.path.dirname(__file__)))
 from fastapi.middleware.cors import CORSMiddleware
 from ktransformers.server.args import ArgumentParser
 from ktransformers.server.config.config import Config
-from ktransformers.server.utils.create_interface import create_interface, GlobalInterface
+from ktransformers.server.utils.create_interface import create_interface, GlobalInterface, get_thread_context_manager
 from fastapi.openapi.utils import get_openapi
 from fastapi import FastAPI
 from fastapi.middleware.cors import CORSMiddleware
@@ -17,6 +17,21 @@ from ktransformers.server.api import router, post_db_creation_operations
 from ktransformers.server.utils.sql_utils import Base, SQLUtil
 from ktransformers.server.config.log import logger
 
+import asyncio
+from uuid import uuid4
+import torch.distributed
+import subprocess
+import tempfile
+import atexit
+
+try:
+    import torch_npu
+    from ktransformers.util import utils
+
+    use_torch_npu = torch_npu.npu.is_available()
+except:
+    use_torch_npu = False
+
 
 def mount_app_routes(mount_app: FastAPI):
     sql_util = SQLUtil()
@@ -100,6 +115,77 @@ def custom_openapi(app):
     return app.openapi_schema
 
 
+def main_npu():
+    torch.npu.config.allow_internal_format = False
+    cfg = Config()
+
+    arg_parser = ArgumentParser(cfg)
+
+    args = arg_parser.parse_args()
+    utils.USE_NPU_GRAPH = args.use_cuda_graph
+    new_chunk_size = min(max(args.chunk_size, 512), utils._MAX_CHUNK_SIZE)
+    if new_chunk_size != args.chunk_size:
+        args.chunk_size = new_chunk_size
+        print(f'[WARN] Chunk size reset to legal value between [512, {utils._MAX_CHUNK_SIZE}] which is {args.chunk_size}.')
+
+    if args.backend_type == "balance_serve":
+        import pickle
+        def cleanup():
+            if sched_process.poll() is None:
+                sched_process.terminate()
+
+        with tempfile.NamedTemporaryFile(delete=False) as temp_file:
+            pickle.dump(args, temp_file)
+            temp_file_path = temp_file.name
+        current_file = __file__
+        target_file = os.path.join(os.path.dirname(current_file), "balance_serve", "sched_rpc.py")
+        target_file = os.path.normpath(target_file)
+        log_path = os.path.join(args.log_dir, "rpc.log")
+        log = open(log_path, "a") 
+        sched_process = subprocess.Popen(
+            ["python3", target_file, "--config", temp_file_path], 
+            stdout=log, 
+            stderr=log
+        )
+        print("sched_rpc started with PID:", sched_process.pid)
+        atexit.register(cleanup)
+    create_interface(config=cfg, default_args=cfg, input_args=args)
+    args.port += torch.distributed.get_rank()
+    tp_size = utils.get_tensor_parallel_size()
+    world_size = torch.distributed.get_world_size()
+    if tp_size == world_size and tp_size > 1:
+        if torch.distributed.get_rank() == 0:
+            app = create_app()
+            custom_openapi(app)
+            run_api(
+                app=app,
+                host=args.host,
+                port=args.port,
+                ssl_keyfile=args.ssl_keyfile,
+                ssl_certfile=args.ssl_certfile,
+            )
+        else:
+            while True:
+                try:
+                    context = get_thread_context_manager()
+                    id = str(uuid4())
+                    context.interface.sync_inference("", id)
+                except Exception as e:
+                    print(f"An error occurred: {e}")
+                finally:
+                    pass
+    else:
+        app = create_app()
+        custom_openapi(app)
+
+        run_api(
+            app=app,
+            host=args.host,
+            port=args.port,
+            ssl_keyfile=args.ssl_keyfile,
+            ssl_certfile=args.ssl_certfile,
+        )
+
 def main():
     cfg = Config()
 
@@ -119,4 +205,7 @@ def main():
     )
 
 if __name__ == "__main__":
-    main()
+    if use_torch_npu:
+        main_npu()
+    else:
+        main()

ktransformers/server/utils/create_interface.py

@@ -16,7 +16,7 @@ from ktransformers.server.backend.interfaces.exllamav2 import ExllamaInterface
 from ktransformers.server.backend.interfaces.transformers import TransformersInterface
 from ktransformers.server.backend.interfaces.ktransformers import KTransformersInterface
 
-def create_interface(config: Config, default_args: ConfigArgs):
+def create_interface(config: Config, default_args: ConfigArgs, input_args=None):
     if config.backend_type=='transformers':
         from ktransformers.server.backend.interfaces.transformers import  TransformersInterface as BackendInterface
     elif config.backend_type == 'exllamav2':
@@ -27,7 +27,12 @@ def create_interface(config: Config, default_args: ConfigArgs):
         from ktransformers.server.backend.interfaces.balance_serve import BalanceServeInterface as BackendInterface
     else:
         raise NotImplementedError(f'{config.backend_type} not implemented')
-    GlobalInterface.interface = BackendInterface(default_args)
+
+    if config.backend_type == 'ktransformers':
+        GlobalInterface.interface = BackendInterface(default_args, input_args)
+    else:
+        GlobalInterface.interface = BackendInterface(default_args)
+
     GlobalContextManager.context_manager = ThreadContextManager(GlobalInterface.interface)
 
 class GlobalContextManager:

ktransformers/util/ascend/ascend_utils.py

@@ -0,0 +1,210 @@
+import os
+from datetime import timedelta
+
+import torch
+try:
+    import torch_npu
+except:
+    Warning("torch_npu not found, please install torch_npu for NPU support.")
+import torch.distributed as dist
+
+_DATA_PARALLEL_SIZE = 0
+_TENSOR_PARALLEL_SIZE = 0
+_DATA_PARALLEL_GROUP = None
+_TENSOR_PARALLEL_RANKS = None
+_TENSOR_PARALLEL_GROUP = None
+_DATA_PARALLEL_GROUP_GLOO = None
+_DATA_PARALLEL_RANKS = None
+_GLOBAL_GROUP = None
+_LM_HEAD_GROUP = None
+
+
+def setup_model_parallel(distributed_timeout_minutes: int = 30, tp: int = 1):
+    global _DATA_PARALLEL_SIZE
+    global _DATA_PARALLEL_GROUP
+    global _DATA_PARALLEL_RANKS
+    global _TENSOR_PARALLEL_SIZE
+    global _TENSOR_PARALLEL_RANKS
+    global _TENSOR_PARALLEL_GROUP
+
+    os.environ["MASTER_ADDR"] = "localhost"
+    local_rank = int(os.getenv("LOCAL_RANK", '0'))
+    world_size = int(os.getenv("WORLD_SIZE", '1'))
+    torch_npu.npu.set_device(local_rank)
+    tp_size = tp
+    dp_size = world_size // tp_size
+    _DATA_PARALLEL_SIZE = dp_size
+    _TENSOR_PARALLEL_SIZE = tp_size
+
+    torch.set_num_threads(8)
+    timeout = timedelta(minutes=distributed_timeout_minutes)
+    print(f"start to init process group ------rank is {local_rank}, world_size is {world_size}")
+    torch.distributed.init_process_group(
+        backend='hccl',
+        world_size=world_size, rank=local_rank
+    )
+    print(f"init process group success ------rank is {local_rank}, world_size is {world_size}")
+
+    rank = torch.distributed.get_rank()
+    nccl_comm_cfgs = {}
+
+    for dp_group_id in range(tp_size):
+        ranks = list(range(dp_group_id, world_size, tp_size))
+        dp_group = torch.distributed.new_group(
+            ranks, timeout=timeout, pg_options=get_nccl_options('dp', nccl_comm_cfgs)
+        )
+        if rank in ranks:
+            global _DATA_PARALLEL_GROUP
+            _DATA_PARALLEL_GROUP = dp_group
+            _DATA_PARALLEL_RANKS = ranks
+
+    for tp_group_id in range(dp_size):
+        start_rank = tp_group_id * tp_size
+        end_rank = (tp_group_id + 1) * tp_size
+        ranks = list(range(start_rank, end_rank))
+        tp_group = torch.distributed.new_group(
+            ranks, timeout=timeout, pg_options=get_nccl_options('tp', nccl_comm_cfgs)
+        )
+        if rank in ranks:
+            global _TENSOR_PARALLEL_GROUP
+            _TENSOR_PARALLEL_GROUP = tp_group
+            _TENSOR_PARALLEL_RANKS = ranks
+
+    torch.manual_seed(1)
+    return local_rank, world_size
+
+
+def get_tensor_parallel_size():
+    assert _TENSOR_PARALLEL_SIZE is not None, "tensor parallel size is not set"
+    return _TENSOR_PARALLEL_SIZE
+
+
+def get_tensor_parallel_group():
+    assert _TENSOR_PARALLEL_GROUP is not None, "tensor parallel group is not initialized"
+    return _TENSOR_PARALLEL_GROUP
+
+
+def get_tensor_parallel_ranks():
+    assert _TENSOR_PARALLEL_RANKS is not None, "tensor parallel ranks is not initialized"
+    return _TENSOR_PARALLEL_RANKS
+
+
+def get_data_parallel_size():
+    assert _DATA_PARALLEL_SIZE is not None, "data parallel size is not initialized"
+    return _DATA_PARALLEL_SIZE
+
+
+def get_data_parallel_gloo():
+    assert _DATA_PARALLEL_GROUP_GLOO is not None, "data parallel gloo group is not initialized"
+    return _DATA_PARALLEL_GROUP_GLOO
+
+
+def get_data_parallel_group():
+    assert _DATA_PARALLEL_GROUP is not None, "data parallel group is not initialized"
+    return _DATA_PARALLEL_GROUP
+
+
+def get_data_parallel_ranks():
+    assert _DATA_PARALLEL_RANKS is not None, "data parallel ranks is not initialized"
+    return _DATA_PARALLEL_RANKS
+
+
+def get_global_group():
+    assert _GLOBAL_GROUP is not None, "global group is not initialized"
+    return _GLOBAL_GROUP
+
+
+def get_nccl_options(pg_name, nccl_comm_cfgs):
+    if pg_name in nccl_comm_cfgs:
+        nccl_options = torch.distributed.ProcessGroupNCCL.Options()
+        nccl_options.config.cga_cluster_size = nccl_comm_cfgs[pg_name].get('cga_cluster_size', 4)
+        nccl_options.config.max_ctas = nccl_comm_cfgs[pg_name].get('max_ctas', 32)
+        nccl_options.config.min_ctas = nccl_comm_cfgs[pg_name].get('min_ctas', 1)
+        return nccl_options
+    else:
+        return None
+
+
+def get_safetensors_cut_weight(name: str, weights: torch.Tensor):
+    translate_col_cut_tensors = ["ffn_down", "attn_output"]
+    translate_row_cut_tensors = ["ffn_gate", "ffn_up", "attn_q_b"]
+    translate_lm_cut_tensor = ["output"]
+    tp = get_tensor_parallel_size()
+    if tp == 1 or weights.shape == torch.Size([1]):
+        return weights
+    world_size = torch.distributed.get_world_size()
+    rank = torch.distributed.get_rank()
+    rank %= tp
+    assert 0 <= rank < tp and tp > 0, f"unexpected {rank=}, {tp=}"
+    if any(t in name for t in translate_col_cut_tensors):
+        if weights.dim() == 1:
+            return weights
+        dim = weights.shape[-1]
+        assert dim % tp == 0, f"unexpected division {dim=}, {tp=}"
+        chunk_size = dim // tp
+        output_weights = weights[:, rank * chunk_size: (rank + 1) * chunk_size]
+        return output_weights
+    elif any(t in name for t in translate_row_cut_tensors):
+        dim = weights.shape[0]
+        assert dim % tp == 0, f"unexpected division {dim=}, {tp=}"
+        chunk_size = dim // tp
+        output_weights = weights[rank * chunk_size: (rank + 1) * chunk_size:]
+        return output_weights
+    elif (tp > 1) and (any(t in name for t in translate_lm_cut_tensor)):
+        dim = weights.shape[0]
+        assert dim % tp == 0, f"unexpected division {dim=} {world_size=}"
+        chunk_size = dim // tp
+        output_weights = weights[rank * chunk_size: (rank + 1) * chunk_size:]
+        return output_weights
+    else:
+        return weights
+
+
+def get_absort_weight(model, config):
+    local_rank = torch.distributed.get_rank()
+    tp = get_tensor_parallel_size()
+    local_rank %= tp
+    tp_heads = config.num_attention_heads // tp
+    for i in range(config.num_hidden_layers):
+        self = model.model.layers[i].self_attn
+        if not (hasattr(self, 'q_absorb') and hasattr(self, 'out_absorb')):
+            kv_b_proj = self.kv_b_proj.weight.view(config.num_attention_heads, -1, self.kv_lora_rank)
+            q_absorb = kv_b_proj[:, :self.qk_nope_head_dim, :].clone()
+            out_absorb = kv_b_proj[:, self.qk_nope_head_dim:, :].clone()
+            q_absorb = q_absorb[local_rank * tp_heads: (local_rank + 1) * tp_heads, :, :].contiguous()
+            out_absorb = out_absorb[local_rank * tp_heads: (local_rank + 1) * tp_heads, :, :].contiguous()
+            out_absorb = out_absorb.transpose(1, 2).contiguous()
+            setattr(self, "q_absorb", q_absorb)
+            setattr(self, "out_absorb", out_absorb)
+            del self.orig_module.kv_b_proj
+
+
+def allreduce_wrapper(func):
+    def wrapper(*args, **kwargs):
+        orig_output = func(*args, **kwargs)
+        if isinstance(orig_output, tuple):
+            if get_tensor_parallel_size() > 1:
+                org_dtype = orig_output[0].dtype
+                if org_dtype == torch.bfloat16:
+                    dist.all_reduce(orig_output[0].to(dtype=torch.float16), op=dist.ReduceOp.SUM,
+                                    group=get_tensor_parallel_group())
+                else:
+                    dist.all_reduce(orig_output[0], op=dist.ReduceOp.SUM, group=get_tensor_parallel_group())
+                if org_dtype == torch.bfloat16:
+                    bf_orig_output = orig_output[0].to(dtype=org_dtype)
+                else:
+                    bf_orig_output = orig_output[0]
+            else:
+                bf_orig_output = orig_output[0]
+            return (bf_orig_output,) + orig_output[1:]
+        else:
+            if get_tensor_parallel_size() > 1:
+                org_dtype = orig_output.dtype
+                if org_dtype == torch.bfloat16:
+                    orig_output = orig_output.to(dtype=torch.float16)
+                dist.all_reduce(orig_output, op=dist.ReduceOp.SUM, group=get_tensor_parallel_group())
+                if org_dtype == torch.bfloat16:
+                    orig_output = orig_output.to(dtype=org_dtype)
+            return orig_output
+
+    return wrapper

ktransformers/util/custom_loader.py

@@ -7,10 +7,19 @@ from typing import Sequence
 import os
 from enum import IntEnum
 import torch
-if not torch.xpu.is_available():
+
+try:
+    import torch_npu
+    use_torch_npu = torch_npu.npu.is_available()
+except:
+    use_torch_npu = False
+
+
+if not torch.xpu.is_available() and not use_torch_npu:
     import KTransformersOps
 from safetensors import safe_open
-from ktransformers.ktransformers_ext.triton.fp8gemm import fp8_gemm, act_quant, weight_dequant
+if not use_torch_npu:
+    from ktransformers.ktransformers_ext.triton.fp8gemm import fp8_gemm, act_quant, weight_dequant
 from ktransformers.util.custom_gguf import *
 from safetensors.torch import save_file
 from abc import ABC, abstractmethod
@@ -42,6 +51,7 @@ class SafeTensorLoader(ModelLoader):
     tensor_device_map: dict
     
     def __init__(self, file_path: str):
+
         self.__load_tensor_file_map(file_path)
 
     def __load_tensor_file_map(self, file_path: str):
@@ -84,6 +94,7 @@ class SafeTensorLoader(ModelLoader):
         # if not found_safetensor:
         #     raise FileNotFoundError(f"No Safetensor files found in {folder_path}")
 
+
     def load_tensor(self, key: str, device: str="cpu"):
         if translate_name_to_gguf(key) in self.tensor_file_map:
             key = translate_name_to_gguf(key)
@@ -96,6 +107,7 @@ class SafeTensorLoader(ModelLoader):
         if f is None:
             raise FileNotFoundError(f"File {file} not found in Safetensor files")
         tensor = f.get_tensor(key)
+      
         return tensor.to(device)
 
     def load_experts(self, key: str, device: str="cpu"):
@@ -252,20 +264,57 @@ class SafeTensorLoader(ModelLoader):
     def has_tensor(self, name: str):
         return name in self.tensor_file_map or translate_name_to_gguf(name) in self.tensor_file_map
 
+
+class W8A8SafeTensorLoader(SafeTensorLoader):
+    def load_tensor(self, key: str, device: str = "cpu"):
+        if key not in self.tensor_file_map:
+            raise KeyError(f"Key {key} not found in Safetensor files")
+        file = self.tensor_file_map[key]
+        f = self.file_handle_map.get(file)
+        if f is None:
+            raise FileNotFoundError(f"File {file} not found in Safetensor files")
+        tensor = f.get_tensor(key)
+        if "deq_scale" in key:
+            tensor = torch.from_numpy(
+                np.frombuffer(tensor.to(torch.float16).to(torch.float32).numpy().tobytes(), dtype=np.int32).astype(np.int64))
+        if "input_scale" in key:
+            tensor = tensor.to(torch.float16)
+        if "weight_scale" in key or "weight_offset" in key:
+            if "ffn" in key:
+                tensor = tensor.to(torch.float32)
+            else:
+                tensor = tensor.to(torch.float16)
+        if "input_offset" in key:
+            tensor = tensor.to(torch.int8)
+        if tensor.dtype == torch.bfloat16:
+            tensor = tensor.to(torch.float16)
+        return tensor.to(device)
+
+    def load_dequantized_tensor(self, key: str, device: str = "cpu"):
+        tensor = self.load_tensor(key, device)
+        return tensor
+
 class GGUFLoader(ModelLoader):
     tensor_info: dict
     gguf_path: str
     tensor_file_map: dict # {tensor_name: tensor_file_path}
     gguf_file_meta: dict
     safetensor_loader: SafeTensorLoader
-    def __init__(self, gguf_path: str):
+    def __init__(self, gguf_path: str, quantize: str = None):
         # Check dir exist
         if not os.path.exists(gguf_path):
             raise FileNotFoundError(f"GGUF dir not found: {gguf_path}")
         if os.path.isfile(gguf_path):
             gguf_path = os.path.dirname(gguf_path)
 
-        self.safetensor_loader = None
+        safetensor_loader = SafeTensorLoader(gguf_path)
+        if quantize == "w8a8_dynamic":
+            safetensor_loader = W8A8SafeTensorLoader(gguf_path)
+        else:
+            safetensor_loader = SafeTensorLoader(gguf_path)
+        if safetensor_loader.tensor_file_map:
+            self.safetensor_loader = safetensor_loader
+            return
         
         self.tensor_info = {}
         self.gguf_path = gguf_path

ktransformers/util/npu_graph.py

@@ -0,0 +1,77 @@
+import time
+
+import torch
+import torch_npu
+import sys
+import os
+
+from ktransformers.util.utils import USE_NPU_GRAPH
+if USE_NPU_GRAPH:
+    CAPTURE_PLUGIN_PATH = os.environ.get("CAPTURE_PLUGIN_PATH")
+    if CAPTURE_PLUGIN_PATH is None:
+        raise RuntimeError("env CAPTURE_PLUGIN_PATH not exist")
+
+    sys.path.append(CAPTURE_PLUGIN_PATH)
+
+    from libgraph_capture import graph_capture_init
+    from libgraph_capture import graph_capture_destroy
+    from libgraph_capture import graph_capture_begin
+    from libgraph_capture import graph_capture_end
+    from libgraph_capture import graph_capture_replay
+    from libgraph_capture import graph_capture_launch_callback
+
+
+class NpuGraph:
+    def init(self):
+        ret = graph_capture_init()
+        if ret != 0:
+            exit()
+
+    def destroy(self):
+        ret = graph_capture_destroy()
+        if ret != 0:
+            exit()
+
+    def capture_begin(
+            self,
+            stream,
+            capture_error_mode="global"):
+        torch.npu.synchronize()
+        torch.npu.empty_cache()
+        ret = graph_capture_begin(stream, capture_error_mode)
+        if ret != 0:
+            exit()
+
+    def capture_end(
+            self,
+            stream):
+        ret = graph_capture_end(stream)
+        if ret != 0:
+            exit()
+
+    def replay(
+            self,
+            stream):
+        ret = graph_capture_replay(stream)
+        if ret != 0:
+            exit()
+
+    def launch_callback(self, func, data, block, stream):
+        graph_capture_launch_callback(func, data, block, stream)
+
+
+class graph:
+    def __init__(
+            self,
+            npu_graph: NpuGraph,
+            pool,
+            stream,
+            capture_error_mode: str = "global"):
+        self.npu_graph = npu_graph
+        self.stream = stream.npu_stream
+
+    def __enter__(self):
+        self.npu_graph.capture_begin(self.stream)
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        self.npu_graph.capture_end(self.stream)

ktransformers/util/npu_graph_runner.py

@@ -0,0 +1,218 @@
+'''
+Description :
+Author      : Boxin Zhang
+Version     : 0.1.0
+Copyright (c) 2024 by KVCache.AI, All Rights Reserved.
+'''
+from typing import Dict
+
+import acl
+import torch
+import torch_npu
+from torch import nn
+
+import ktransformers.util.npu_graph as npu_graph
+from ktransformers.util.utils import CUR_DEVICE
+
+
+class NPUGraphRunner:
+    def __init__(self, deviceId):
+        torch.npu.set_compile_mode(jit_compile=False)
+        self.deviceId = deviceId
+        self.enable = False
+        self.debug = False
+        self.input_buffers: Dict[str, torch.Tensor] = {}
+        self.output_buffers: Dict[str, torch.Tensor] = {}
+        self.tid = None
+        self.past_key_value = None
+
+    def init(self, batch_size, seq_length):
+        self.tmp_g = npu_graph.NpuGraph()
+        self.graph = torch.npu.NPUGraph()
+        self.main_stream = torch_npu.npu.Stream(device=self.deviceId)
+        self.update_stream = torch_npu.npu.Stream(device=self.deviceId)
+        self.stream = self.main_stream.npu_stream
+        self.logits = torch.zeros((batch_size, seq_length, 7168), dtype=torch.float16).to(self.deviceId)
+        self.context, ret = acl.rt.get_context(self.deviceId)
+        if ret != 0:
+            print("get_context failed! ret: " + str(ret))
+            exit(-1)
+        self.exit_flag = False
+        self.handle = []
+        self.ifa_param = []
+        self.event = []
+        self.first_update = True
+        self.workspace = None
+
+        if self.tid is None:
+            def process_callback(args_list):
+                ins = args_list[0]
+                ret = acl.rt.set_context(ins.context)
+                if ret != 0:
+                    print("set_context failed! ret: " + str(ret))
+                    exit(-1)
+
+                while True:
+                    acl.rt.process_report(1)
+                    if ins.exit_flag:
+                        break
+
+            self.tid, ret = acl.util.start_thread(process_callback, [self])
+            if ret != 0:
+                print("start_thread failed!")
+                exit(-1)
+
+        ret = acl.rt.subscribe_report(self.tid, self.stream)
+        if ret != 0:
+            print("subscribe_report failed!")
+            exit(-1)
+
+    def destroy(self):
+        print(f'[rank:{torch.distributed.get_rank()}]------------- NPU Graph Destroy Begin -------------\n', end='')
+        self.exit_flag = True
+        ret = acl.rt.unsubscribe_report(self.tid, self.stream)
+        if ret != 0:
+            print("unsubscribe_report failed!")
+            exit(-1)
+        self.enable = False
+        ret = acl.util.stop_thread(self.tid)
+        if ret != 0:
+            print("stop_thread failed!")
+            exit(-1)
+        self.tid = None
+        self.workspace = None
+        self.handle = []
+        self.ifa_param = []
+        self.event = []
+        self.first_update = True
+        del self.graph
+        self.tmp_g.destroy()
+        destroy_runner(self.deviceId)
+        print(f'[rank:{torch.distributed.get_rank()}]------------- NPU Graph Destroy Finish -------------\n', end='')
+
+    def capture(
+            self,
+            model,
+            cur_token,
+            position_ids,
+            cache_position,
+            past_key_values,
+            main_device,
+            **kwargs,
+    ) -> None:
+        print(f'[rank:{torch.distributed.get_rank()}]------------- NPU Graph Capture Begin -------------\n', end='')
+        self.enable = True
+        self.model = model
+        inputs_embeds = model.model.embed_tokens(cur_token.to("cpu")).to(main_device)
+        self.seq_length = inputs_embeds.size()[1]
+        self.main_device = main_device
+        with torch.no_grad():
+            with torch.npu.graph(self.graph, stream=self.main_stream):
+                self.logits = model(inputs_embeds=inputs_embeds,
+                                    position_ids=position_ids,
+                                    cache_position=cache_position,
+                                    past_key_values=past_key_values,
+                                    **kwargs)[0]
+
+        if past_key_values != None:
+            past_key_values.change_seq_length(-1)
+
+        self.input_buffers = {
+            "inputs_embeds": inputs_embeds,
+            "position_ids": position_ids,
+            "cache_position": cache_position,
+        }
+
+        self.output_buffers = {"logits": self.logits}
+        print(f'[rank:{torch.distributed.get_rank()}]------------- NPU Graph Capture Finish -------------\n', end='')
+        return
+
+    def forward(
+            self,
+            inputs_embeds,
+            position_ids,
+            cache_position,
+    ) -> torch.Tensor:
+        def ifa_update_sync(param):
+            with torch.npu.stream(self.update_stream):
+                for i in range(len(self.handle)):
+                    if self.first_update is False:
+                        q_nope, kvCache, q_pe, kRopeCache, num_heads, \
+                            softmax_scale, layer_idx, attn_output, softmax_lse = self.ifa_param[i]
+                        torch.npu.graph_task_update_begin(self.update_stream, self.handle[i])
+                        torch_npu.npu_fused_infer_attention_score.out(
+                            q_nope,
+                            kvCache,
+                            kvCache,
+                            workspace=self.workspace,
+                            query_rope=q_pe,
+                            key_rope=kRopeCache,
+                            num_heads=num_heads,
+                            num_key_value_heads=1,
+                            input_layout="BNSD",
+                            atten_mask=None,
+                            scale=softmax_scale,
+                            antiquant_mode=0,
+                            antiquant_scale=None,
+                            block_table=self.past_key_value.page_table_list[layer_idx],
+                            block_size=self.past_key_value.page_size,
+                            actual_seq_lengths_kv=self.past_key_value.position,
+                            out=[attn_output, softmax_lse])
+                        torch.npu.graph_task_update_end(self.update_stream)
+                    self.event[i].record(self.update_stream)
+
+        self.ifa_update_tid, ret = acl.util.start_thread(ifa_update_sync, [self])
+        if ret != 0:
+            print("start_thread failed!")
+            exit(-1)
+
+        ret1 = acl.rt.memcpy(self.input_buffers["inputs_embeds"].data_ptr(), inputs_embeds.numel() * 2,
+                             inputs_embeds.data_ptr(), inputs_embeds.numel() * 2, 3)
+        ret2 = acl.rt.memcpy(self.input_buffers["position_ids"].data_ptr(), position_ids.numel() * 8,
+                             position_ids.data_ptr(), position_ids.numel() * 8, 3)
+        ret3 = acl.rt.memcpy(self.input_buffers["cache_position"].data_ptr(), cache_position.numel() * 8,
+                             cache_position.data_ptr(), cache_position.numel() * 8, 3)
+        torch_npu.npu.synchronize()
+
+        with torch_npu.npu.stream(self.main_stream):
+            self.graph.replay()
+        self.first_update = False
+        ret = acl.util.stop_thread(self.ifa_update_tid)
+        if ret != 0:
+            print("stop_thread failed!")
+            exit(-1)
+        else:
+            self.ifa_update_tid = None
+        return self.output_buffers["logits"]
+
+    def launch_callback(self, func, data, block, stream):
+        self.tmp_g.launch_callback(func, data, block, stream)
+
+    def __call__(self, *args, **kwargs):
+        return self.forward(*args, **kwargs)
+
+
+runner_dict = dict()
+
+
+def check_runner(deviceId: int):
+    runner = runner_dict.get(deviceId)
+    if runner is None:
+        return True
+    else:
+        return False
+
+
+def destroy_runner(deviceId: int):
+    runner = runner_dict.get(deviceId)
+    if runner is not None:
+        runner_dict[deviceId] = None
+
+
+def get_or_create_runner(deviceId: int):
+    runner = runner_dict.get(deviceId)
+
+    if runner is None:
+        runner = NPUGraphRunner(deviceId)
+        runner_dict[deviceId] = runner
+    return runner

ktransformers/util/utils.py

@@ -31,8 +31,35 @@ if not torch.xpu.is_available():
     from ktransformers.operators.flashinfer_wrapper import MLAWrapperSingleton
 import socket
 
+import os
+import re
+import torch.distributed as dist
+try:
+    import torch_npu
+    from ktransformers.util.ascend.ascend_utils import get_tensor_parallel_size
+    use_torch_npu = torch_npu.npu.is_available()
+except:
+    use_torch_npu = False
+
+
 warm_uped = False
 
+
+W8A8_ENABLE = False
+Q4_GGUF_LODER = None
+USE_NPU_GRAPH = None
+WARM_UP_SKIP_CNT = [1, 1]
+_USE_NPU_GRAPH = False
+_MAX_DECODE_PROFILE = 3
+CUR_DEVICE = None
+_MAX_CHUNK_SIZE = int(max(os.getenv("_MAX_CHUNK_SIZE", 4096), 512))
+
+
+def get_use_npu_graph():
+    assert _USE_NPU_GRAPH is not None, "use npu graph is not setting"
+    return _USE_NPU_GRAPH
+
+
 def get_free_ports(n: int, continue_prot: list):
     sockets = []
     ports = []
@@ -50,6 +77,10 @@ def get_free_ports(n: int, continue_prot: list):
     return ports
 
 def get_compute_capability(device:torch.device = None):
+
+    if use_torch_npu:
+        return 0
+
     if torch.cuda.is_available():
         if device is None:
             num_gpus = torch.cuda.device_count()
@@ -97,9 +128,16 @@ def get_all_used_cuda_device(device_map:dict):
         all_device_list.add(device_map[key]["prefill_device"]) if "prefill_device" in device_map[key] else None
     if "cpu" in all_device_list:
         all_device_list.remove("cpu")
+
+    if use_torch_npu:
+        all_device_list = set([device.replace("cuda", "npu") for device in all_device_list])
+
     all_device_list = list(all_device_list)
     return all_device_list
 
+
+
+# TODO: support NPU
 def load_cur_state_dict(module: nn.Module, gguf_loader: ModelLoader, prefix: str = "", device="cuda"):
     prefix = prefix.replace("orig_module.", "")
     persistent_buffers = {k: v for k, v in module._buffers.items() if k not in module._non_persistent_buffers_set}
@@ -109,6 +147,7 @@ def load_cur_state_dict(module: nn.Module, gguf_loader: ModelLoader, prefix: str
         key = prefix + name
         translated_key = key
         
+
         # TODO: Merge all loader.
         # I know this is ugly but lets do it for now.
         if isinstance(gguf_loader, SafeTensorLoader):
@@ -120,7 +159,13 @@ def load_cur_state_dict(module: nn.Module, gguf_loader: ModelLoader, prefix: str
         if gguf_loader.has_tensor(translated_key) or "kv_b_proj" in translated_key:
             target_dtype = torch.get_default_dtype()
             device = get_device(translated_key[:translated_key.rfind(".")], gguf_loader.tensor_device_map)
-            print(f"loading {translated_key} to {device}")
+
+
+            if use_torch_npu:
+                device = "cpu" if "embd" in translated_key else CUR_DEVICE
+                print(f"loading layer {translated_key} to {device}") if torch.distributed.get_rank() == 0 else None
+            else:
+                print(f"loading {translated_key} to {device}")
             if torch.cuda.is_available():
                 torch.cuda.empty_cache()
             elif torch.xpu.is_available():
@@ -149,6 +194,8 @@ def sync_all_device(all_device_list):
             torch.cuda.synchronize(device)
         elif "xpu" in device.lower():
             torch.xpu.synchronize(device)
+        elif use_torch_npu:
+            torch_npu.synchronize(device)
         else:
             raise RuntimeError("The device {} is not available".format(device))
 
@@ -228,20 +275,68 @@ def tf_logits_warper(generation_config):
 
 def prefill_and_generate(model, tokenizer, inputs, max_new_tokens=10000, use_cuda_graph: bool = True,
                          mode = 'normal', force_think: bool = False, chunk_size = 16384, use_flashinfer_mla = False,
-                         num_heads = None, head_dim_ckv = None, head_dim_kpe = None, q_head_dim = None):
+                         num_heads = None, head_dim_ckv = None, head_dim_kpe = None, q_head_dim = None, static_cache = None):
     import os
     os.environ["TOKENIZERS_PARALLELISM"] = "false"
     torch._dynamo.config.suppress_errors = True
     batch_size, seq_length = inputs.shape
     device_map = model.gguf_loader.tensor_device_map
-    torch_device = get_device('model.layers.0.self_attn', device_map)
-    torch_device = torch_device_mapping[torch_device] if torch_device in torch_device_mapping else torch_device
+
+    if use_torch_npu:
+        vocabulary_size = model.config.vocab_size
+        topp = torch.tensor([[model.generation_config.top_p]], dtype=torch.float16).npu()
+        topk = torch.tensor([[model.generation_config.top_k]], dtype=torch.int32).npu()
+        temperature = torch.tensor([[model.generation_config.temperature]], dtype=torch.float16).npu()
+        next_token_fake = torch.tensor([[1]], dtype=torch.int32).npu()
+        next_token_probs = torch.tensor([[1.0]], dtype=torch.float16).npu()
+        torch_device = CUR_DEVICE
+    else:
+        torch_device = get_device('model.layers.0.self_attn', device_map)
+        torch_device = torch_device_mapping[torch_device] if torch_device in torch_device_mapping else torch_device
     inputs = inputs.to(torch_device)
     all_cuda_device = get_all_used_cuda_device(device_map)
 
     tokens = []
     
+    def decode_one_tokens_npu(cuda_graph_runner, cur_token, position_ids, cache_position, past_key_values, logits_warper, generation_config, use_cuda_graph: bool = True):
+        if cuda_graph_runner is None:
+            use_cuda_graph = False
+        inputs_embeds = model.model.embed_tokens(cur_token.to('cpu')).to(torch_device)
+        if use_cuda_graph:
+            logits = cuda_graph_runner(inputs_embeds, position_ids, cache_position)
+        else:
+            # custom_stream = torch.cuda.Stream()
+            # torch.cuda.set_device(torch_device)
+            torch_npu.npu.set_device(torch_device)
+            # with torch.cuda.stream(custom_stream):
+            logits=model(inputs_embeds=inputs_embeds,
+                        position_ids=position_ids,
+                        cache_position=cache_position,
+                        past_key_values=past_key_values,
+                        return_dict=False, use_cache=True)[0]
+        if past_key_values != None:
+            past_key_values.change_seq_length(1)
+        all_cuda_device = ['npu:' + str(index) for index in range(torch.distributed.get_world_size())]
+        for device in all_cuda_device:
+            # torch.cuda.synchronize(device)
+            torch_npu.npu.synchronize(device)
+        if generation_config.do_sample:
+            logits = logits / temperature
+            torch.manual_seed(0)
+            probs = logits.view(batch_size, vocabulary_size)
+            sm = nn.Softmax(dim=-1)
+            probs = sm(probs).half().npu()
+            next_token = next_token_fake
+            torch_npu._npu_topk_topp_sampling(probs, topk, topp, next_token, next_token_probs)
+            next_token = next_token.squeeze(-1)
+        else:
+            next_token_scores = logits_warper(inputs, logits[:, -1, :])
+            next_token = torch.argmax(next_token_scores, dim=-1)
+        return next_token
+
     def decode_one_tokens(cuda_graph_runner, cur_token, position_ids, cache_position, past_key_values, logits_warper, generation_config, use_cuda_graph: bool = True):
+        if use_torch_npu:
+            return decode_one_tokens_npu(cuda_graph_runner, cur_token, position_ids, cache_position, past_key_values, logits_warper, generation_config, use_cuda_graph)
         if cuda_graph_runner is None:
             use_cuda_graph = False
         if use_cuda_graph:
@@ -252,6 +347,8 @@ def prefill_and_generate(model, tokenizer, inputs, max_new_tokens=10000, use_cud
                 torch.cuda.set_device(torch_device)
             elif torch.xpu.is_available():
                 torch.xpu.set_device(torch_device)
+            elif use_torch_npu:
+                torch_npu.set_device(torch_device)
             else:
                 raise RuntimeError(f"The device: {torch_device} is not available")
             inputs_embeds = model.model.embed_tokens(cur_token.to("cpu")).to(torch_device)
@@ -279,6 +376,7 @@ def prefill_and_generate(model, tokenizer, inputs, max_new_tokens=10000, use_cud
             inputs_embeds = model.model.embed_tokens(inputs.to("cpu"))
         else:
             inputs_embeds = model.model.embed_tokens(inputs.to("cpu")).to(torch_device)
+
         if use_flashinfer_mla:
             MLAWrapperSingleton.update_buffer(past_key_values.max_pages)
             MLAWrapperSingleton.need_plan_all()
@@ -288,11 +386,88 @@ def prefill_and_generate(model, tokenizer, inputs, max_new_tokens=10000, use_cud
         )[0][:,-1,:].unsqueeze(0).clone().to(torch_device)
         
         return logits
-    
+
+    def decode_wrapper(next_token, position_ids, cache_position, cuda_graph_runner, past_key_values, inputs, seq_length, prof=None):
+        global warm_uped
+        global _USE_NPU_GRAPH
+        if use_cuda_graph:
+            from ktransformers.util.npu_graph_runner import get_or_create_runner
+            npu_graph_runner = get_or_create_runner(CUR_DEVICE)
+            npu_graph_runner.init(batch_size, seq_length)
+            with torch_npu.npu.stream(npu_graph_runner.main_stream):
+                for i in range(1, max_new_tokens):
+                    if use_flashinfer_mla:
+                        MLAWrapperSingleton.plan_all(None, None, None, position_ids.squeeze(1) + 1, None,
+                                                     num_heads, head_dim_ckv, head_dim_kpe, past_key_values.page_size,
+                                                     model.model.layers[0].self_attn.softmax_scale, torch.bfloat16,
+                                                     torch.bfloat16)
+                    if use_cuda_graph and ((warm_uped == True and int(i) == 1) or (warm_uped == False and int(i) == 2)):
+                        warm_uped = True
+                        _USE_NPU_GRAPH = True
+                        npu_graph_runner.capture(model, next_token.unsqueeze(0), position_ids, cache_position, past_key_values, torch_device, return_dict=False, use_cache=True)
+                        cuda_graph_runner = npu_graph_runner
+
+                    next_token = decode_one_tokens(cuda_graph_runner, next_token.unsqueeze(0), position_ids,
+                                                   cache_position, past_key_values, logits_warper, generation_config,
+                                                   use_cuda_graph).to(torch_device)
+                    inputs = torch.cat((inputs, next_token.unsqueeze(0)), dim=-1)
+                    generated_ids[:, cache_position] = next_token.int()
+                    tokens.append(int(next_token))
+                    seq_length += 1
+
+                    if next_token[0].item() == tokenizer.eos_token_id or tokenizer.decode(
+                            next_token.tolist()) == '<|im_end|>':
+                        print(stream.end(), end="", flush=True)
+                        break
+                    else:
+                        if torch.distributed.get_rank() % get_tensor_parallel_size() == 0:
+                            print(stream.put(next_token.item()), end="", flush=True)
+                    cache_position += 1
+                    past_key_values.position[0] += 1
+                    position_ids = cache_position.unsqueeze(0)
+
+                    if prof is not None:
+                        prof.step()
+                npu_graph_runner.destroy()
+                _USE_NPU_GRAPH = False
+        else:
+            for i in range(1, max_new_tokens):
+                if use_flashinfer_mla:
+                    MLAWrapperSingleton.plan_all(None, None, None, position_ids.squeeze(1) + 1, None,
+                                                 num_heads, head_dim_ckv, head_dim_kpe, past_key_values.page_size,
+                                                 model.model.layers[0].self_attn.softmax_scale, torch.bfloat16,
+                                                 torch.bfloat16)
+                next_token = decode_one_tokens(cuda_graph_runner, next_token.unsqueeze(0), position_ids, cache_position,
+                                               past_key_values, logits_warper, generation_config, use_cuda_graph).to(
+                    torch_device)
+                inputs = torch.cat((inputs, next_token.unsqueeze(0)), dim=-1)
+                generated_ids[:, cache_position] = next_token.int()
+                tokens.append(int(next_token))
+                seq_length += 1
+
+                if next_token[0].item() == tokenizer.eos_token_id or tokenizer.decode(
+                        next_token.tolist()) == '<|im_end|>':
+                    print(stream.end(), end="", flush=True)
+                    break
+                else:
+                    if torch.distributed.get_rank() % get_tensor_parallel_size() == 0:
+                        print(stream.put(next_token.item()), end="", flush=True)
+                cache_position += 1
+                past_key_values.position[0] += 1
+                position_ids = cache_position.unsqueeze(0)
+
+                if prof is not None:
+                    prof.step()
+        if prof is not None:
+            prof.stop()
+
+
     if torch.cuda.is_available():
         torch.cuda.set_device(torch_device)
     elif torch.xpu.is_available():
         torch.xpu.set_device(torch_device)
+    elif use_torch_npu:
+        torch_npu.set_device(torch_device)
     else:
         raise RuntimeError(f"The device: {torch_device} is not available")
     with torch.no_grad():
@@ -304,6 +479,16 @@ def prefill_and_generate(model, tokenizer, inputs, max_new_tokens=10000, use_cud
                 past_key_values = DynamicUnbalancedFp8Cache.from_legacy_cache(None)
             else:
                 past_key_values = DynamicNormalCache.from_legacy_cache(None)
+        elif use_torch_npu and static_cache:
+            assert isinstance(static_cache, StaticCache), '[ERROR] static_cache format not equal to StaticCache'
+            past_key_values = static_cache
+            if past_key_values.max_batch_size < batch_size or past_key_values.max_cache_len < seq_length + max_new_tokens:
+                print('[WARN] current staticCache size exceeded, try create new staticCache...')
+                past_key_values = StaticCache(
+                    config=model.config, max_batch_size=1, max_cache_len=seq_length + max_new_tokens, device=device_map, dtype=model.dtype
+                )
+            else:
+                past_key_values.reset()
         elif mode != 'long_context':
             past_key_values = StaticCache(
                 config = model.config, max_batch_size = 1, max_cache_len = seq_length + max_new_tokens, device = device_map, dtype = model.dtype
@@ -320,19 +505,67 @@ def prefill_and_generate(model, tokenizer, inputs, max_new_tokens=10000, use_cud
         logits_warper = tf_logits_warper(generation_config)
 
         cache_position = torch.arange(seq_length, device=torch_device, dtype=torch.int32)
+        if use_torch_npu:
+            past_key_values.position[0] = seq_length + 1
+
         generated_ids = torch.zeros(
             batch_size, seq_length + max_new_tokens + 1, dtype=torch.int, device=torch_device
         )
         generated_ids[:, cache_position] = inputs.to(torch_device).to(torch.int)
         start_time = time.time()
 
-        chunk_start = 0
-        while chunk_start < seq_length:
-            chunk_end = min(chunk_start + chunk_size, seq_length)
-            if past_key_values != None:
-                past_key_values.cur_idx=cache_position[chunk_start:chunk_end]
-            logits = chunk_prefill(inputs[:, chunk_start:chunk_end], cache_position[chunk_start:chunk_end], past_key_values)
-            chunk_start += chunk_size
+        logits = None
+
+        def prefill_wrapper(prof=None):
+            nonlocal logits
+            chunk_start = 0
+            while chunk_start < seq_length:
+                chunk_end = min(chunk_start + chunk_size, seq_length)
+                if past_key_values != None:
+                    past_key_values.cur_idx=cache_position[chunk_start:chunk_end]
+                logits = chunk_prefill(inputs[:, chunk_start:chunk_end], cache_position[chunk_start:chunk_end], past_key_values)
+                chunk_start += chunk_size
+                if prof is not None:
+                    prof.step()
+            if prof is not None:
+                prof.stop()
+            if logits is None:
+                raise ValueError('logits cannot be None')
+
+        if use_torch_npu:
+            global WARM_UP_SKIP_CNT
+            prof_prefill = os.environ["PROF_PREFILL"] if "PROF_PREFILL" in os.environ else "0"
+            if prof_prefill == "1" and WARM_UP_SKIP_CNT[0] <= 0:
+                experimental_config = torch_npu.profiler._ExperimentalConfig(
+                    aic_metrics=torch_npu.profiler.AiCMetrics.PipeUtilization,
+                    profiler_level=torch_npu.profiler.ProfilerLevel.Level1, l2_cache=False
+                )
+                with torch_npu.profiler.profile(
+                        activities=[
+                            torch_npu.profiler.ProfilerActivity.CPU,
+                            torch_npu.profiler.ProfilerActivity.NPU
+                        ],
+                        schedule=torch_npu.profiler.schedule(wait=0, warmup=0, active=8, repeat=1, skip_first=0),
+                        on_trace_ready=torch_npu.profiler.tensorboard_trace_handler("./prefill_prof"),
+                        record_shapes=True,
+                        profile_memory=True,
+                        with_stack=False,
+                        with_flops=False,
+                        with_modules=False,
+                        experimental_config=experimental_config) as prof:
+                    prefill_wrapper(prof)
+            else:
+                prefill_wrapper()
+            WARM_UP_SKIP_CNT[0] -= 1
+        else:
+
+            chunk_start = 0
+            while chunk_start < seq_length:
+                chunk_end = min(chunk_start + chunk_size, seq_length)
+                if past_key_values != None:
+                    past_key_values.cur_idx=cache_position[chunk_start:chunk_end]
+                logits = chunk_prefill(inputs[:, chunk_start:chunk_end], cache_position[chunk_start:chunk_end], past_key_values)
+                chunk_start += chunk_size
 
         next_token_scores = logits_warper(inputs, logits[:, -1, :])
         if generation_config.do_sample:
@@ -348,56 +581,106 @@ def prefill_and_generate(model, tokenizer, inputs, max_new_tokens=10000, use_cud
 
         prefill_count = seq_length
         prefill_time = first_token_time
-        if force_think:
-            print("<think>")
-        print(stream.put(next_token.item()), end="", flush=True)
+        if use_torch_npu and torch.distributed.get_rank() % get_tensor_parallel_size() == 0:
+            if force_think:
+                print("<think>")
+            print(stream.put(next_token.item()), end="", flush=True)
+        elif not use_torch_npu:
+            if force_think:
+                print("<think>")
+            print(stream.put(next_token.item()), end="", flush=True)            
+
         generated_ids[:, seq_length] = next_token
         tokens.append(int(next_token))
         inputs = torch.cat((inputs, next_token.unsqueeze(0)), dim=-1)
         cache_position = torch.tensor([seq_length], device=torch_device, dtype=torch.int32)
         position_ids = cache_position.unsqueeze(0)
         seq_length += 1
+        if use_torch_npu:
+            past_key_values.position += 1
         
         cuda_graph_runner = None
             
         start_time = time.time()
-        for i in range(1, max_new_tokens):
-            if use_flashinfer_mla:
-                MLAWrapperSingleton.plan_all(None,None,None,position_ids.squeeze(1)+1,None,
-                                             num_heads, head_dim_ckv, head_dim_kpe, past_key_values.page_size,
-                                             model.model.layers[0].self_attn.softmax_scale, torch.bfloat16, torch.bfloat16)
-            global warm_uped
-            if use_cuda_graph and ( (warm_uped == True and int(i) == 1) or (warm_uped == False and int(i) == 2) ):
-                warm_uped = True
-                cuda_graph_runner = CUDAGraphRunner()
-                cuda_graph_runner.capture(model, next_token.unsqueeze(0), position_ids, cache_position, past_key_values, torch_device, return_dict=False, use_cache=True)
-            next_token = decode_one_tokens(cuda_graph_runner, next_token.unsqueeze(0), position_ids, cache_position, past_key_values, logits_warper, generation_config, use_cuda_graph).to(torch_device)
-            inputs = torch.cat((inputs, next_token.unsqueeze(0)), dim=-1)
-            generated_ids[:, cache_position] = next_token.int()
-            tokens.append(int(next_token))
-            seq_length += 1
-            
-            if next_token[0].item() == tokenizer.eos_token_id or tokenizer.decode(next_token.tolist()) == '<|im_end|>':
-                print(stream.end(), end="", flush=True)
-                break
+
+        if not use_torch_npu:
+            for i in range(1, max_new_tokens):
+                if use_flashinfer_mla:
+                    MLAWrapperSingleton.plan_all(None,None,None,position_ids.squeeze(1)+1,None,
+                                                num_heads, head_dim_ckv, head_dim_kpe, past_key_values.page_size,
+                                                model.model.layers[0].self_attn.softmax_scale, torch.bfloat16, torch.bfloat16)
+                global warm_uped
+                if use_cuda_graph and ( (warm_uped == True and int(i) == 1) or (warm_uped == False and int(i) == 2) ):
+                    warm_uped = True
+                    cuda_graph_runner = CUDAGraphRunner()
+                    cuda_graph_runner.capture(model, next_token.unsqueeze(0), position_ids, cache_position, past_key_values, torch_device, return_dict=False, use_cache=True)
+                next_token = decode_one_tokens(cuda_graph_runner, next_token.unsqueeze(0), position_ids, cache_position, past_key_values, logits_warper, generation_config, use_cuda_graph).to(torch_device)
+                inputs = torch.cat((inputs, next_token.unsqueeze(0)), dim=-1)
+                generated_ids[:, cache_position] = next_token.int()
+                tokens.append(int(next_token))
+                seq_length += 1
+                
+                if next_token[0].item() == tokenizer.eos_token_id or tokenizer.decode(next_token.tolist()) == '<|im_end|>':
+                    print(stream.end(), end="", flush=True)
+                    break
+                else:
+                    print(stream.put(next_token.item()), end="", flush=True)
+                cache_position += 1
+                position_ids = cache_position.unsqueeze(0)
+        else:
+            prof_decode = os.environ["PROF_DECODE"] if "PROF_DECODE" in os.environ else "0"
+            prof_ranks = os.environ["PROF_RANK"] if "PROF_RANK" in os.environ else "0"
+            prof_ranks = [int(r.strip()) for r in prof_ranks.split(",")]
+            if prof_decode == "1" and torch.distributed.get_rank() in prof_ranks and WARM_UP_SKIP_CNT[1] <= 0:
+                experimental_config = torch_npu.profiler._ExperimentalConfig(
+                    aic_metrics=torch_npu.profiler.AiCMetrics.PipeUtilization,
+                    profiler_level=torch_npu.profiler.ProfilerLevel.Level1, l2_cache=False
+                )
+                with torch_npu.profiler.profile(
+                        activities=[
+                            torch_npu.profiler.ProfilerActivity.CPU,
+                            torch_npu.profiler.ProfilerActivity.NPU
+                        ],
+                        schedule=torch_npu.profiler.schedule(wait=0, warmup=0, active=_MAX_DECODE_PROFILE, repeat=1, skip_first=0),
+                        on_trace_ready=torch_npu.profiler.tensorboard_trace_handler("./decode_prof"),
+                        record_shapes=True,
+                        profile_memory=True,
+                        with_stack=False,
+                        with_flops=False,
+                        with_modules=False,
+                        experimental_config=experimental_config) as prof:
+                    decode_wrapper(next_token, position_ids, cache_position, cuda_graph_runner, past_key_values, inputs, seq_length, prof)
             else:
-                print(stream.put(next_token.item()), end="", flush=True)
-            cache_position += 1
-            position_ids = cache_position.unsqueeze(0)
+                decode_wrapper(next_token, position_ids, cache_position, cuda_graph_runner, past_key_values, inputs, seq_length)
+            WARM_UP_SKIP_CNT[1] -= 1            
         
 
     total_time = time.time() - start_time
     tokens_generated = len(tokens)
     tokens_per_second = tokens_generated / total_time
 
-    print("")
+    if not use_torch_npu:
+        print("")
+
+        print(f"prompt eval count:    {prefill_count} token(s)")
+        print(f"prompt eval duration: {prefill_time}s")
+        print(f"prompt eval rate:     {prefill_count/prefill_time} tokens/s")
+        print(f"eval count:           {tokens_generated} token(s)")
+        print(f"eval duration:        {total_time}s")
+        print(f"eval rate:            {tokens_per_second} tokens/s")
+    else:
+        tp_size = get_tensor_parallel_size()
+        if torch.distributed.get_rank() % tp_size == 0:
+            rank = f"[rank:{torch.distributed.get_rank()}]"
+            msg = f"\n{rank} Eval Time\n"
+            msg += rank + f"prompt eval count:    {prefill_count} token(s)\n"
+            msg += rank + f"prompt eval duration: {prefill_time:.9f}s\n"
+            msg += rank + f"prompt eval rate:     {prefill_count/prefill_time:.9f} tokens/s\n"
+            msg += rank + f"eval count:           {tokens_generated} token(s)\n"
+            msg += rank + f"eval duration:        {total_time:.9f}s\n"
+            msg += rank + f"eval rate:            {tokens_per_second:.9f} tokens/s\n"
+            print(msg)
 
-    print(f"prompt eval count:    {prefill_count} token(s)")
-    print(f"prompt eval duration: {prefill_time}s")
-    print(f"prompt eval rate:     {prefill_count/prefill_time} tokens/s")
-    print(f"eval count:           {tokens_generated} token(s)")
-    print(f"eval duration:        {total_time}s")
-    print(f"eval rate:            {tokens_per_second} tokens/s")
 
     return tokens
 

merge_tensors/merge_safetensor_gguf.py

@@ -12,6 +12,8 @@ from safetensors.torch import save_file
 import re
 from collections import defaultdict
 
+SKIP_MTP = True
+
 def read_safetensor_keys_from_folder(folder_path)->dict:
     """    
     :param folder_path: folder path
@@ -36,7 +38,7 @@ def read_safetensor_keys_from_folder(folder_path)->dict:
                 try:
                     with safe_open(file_path, framework="pt") as f:
                         for key in f.keys():
-                            if "model.layers.61" in key:
+                            if SKIP_MTP and "model.layers.61" in key:
                                 # skip MTP layer
                                 continue
                             # try:
@@ -94,6 +96,28 @@ def combine_tensor_sources(safetensor_path:str, gguf_path:str):
     
     return target_tensor_map, gguf_loader
 
+def combine_w8a8_tensor_sources(safetensor_path: str, gguf_path: str):
+    gguf_loader = GGUFLoader(gguf_path)
+    gguf_tensor_file_map = gguf_loader.tensor_file_map
+    safetensor_tensor_file_map = read_safetensor_keys_from_folder(safetensor_path)
+
+    # build a map for the key to the tensor
+    # according to the key, we can get the tensor from the file
+
+    target_tensor_map = {}
+    for key in safetensor_tensor_file_map.keys():
+        # for all experts, we use the gguf tensor
+        if ".mlp.experts." in key and "weight_scale" not in key and "weight_offset" not in key:
+            key = '.'.join(key.split('.')[:5] + key.split('.')[-2:])
+            translated_key = translate_name(key)
+            target_tensor_map[key] = gguf_tensor_file_map[translated_key]
+        elif ".mlp.experts." in key and ("weight_scale" not in key or "weight_offset" not in key):
+            continue
+        else:
+            target_tensor_map[key] = safetensor_tensor_file_map[key]
+
+    return target_tensor_map, gguf_loader
+
 def write_combined_tensor(target_tensor_map: dict, output_path: str, gguf_loader: GGUFLoader):
     # Ensure output directory exists
     os.makedirs(output_path, exist_ok=True)
@@ -193,6 +217,7 @@ def main():
     parser.add_argument("--safetensor_path", type=str, help="Path to the Safetensor file", default="/mnt/data/model/DeepSeek-V3")
     parser.add_argument("--gguf_path", type=str, help="Path to the GGUF file", default="/mnt/data/model/DeepseekV3-q4km-gguf")
     parser.add_argument("--output_path", type=str, help="Path to the output file", default="/mnt/data/model/ktrans-safetensors/DeepSeek-V3-q4km-fp8")
+    parser.add_argument("--safetensors_format", type=str, help="Safetensors format", default="fp8")
     
     # print all the arguments
     print("All the arguments:")
@@ -204,8 +229,18 @@ def main():
     safetensor_path = args.safetensor_path
     gguf_path = args.gguf_path
     output_path = args.output_path
+    safetensors_format = args.safetensors_format
     
-    target_tensor_map, gguf_loader = combine_tensor_sources(safetensor_path, gguf_path)
+    match safetensors_format:
+        case "w8a8":
+            global SKIP_MTP
+            SKIP_MTP = False
+            target_tensor_map, gguf_loader = combine_w8a8_tensor_sources(safetensor_path, gguf_path)
+        case "fp8":
+            target_tensor_map, gguf_loader = combine_tensor_sources(safetensor_path, gguf_path)
+        case _:
+            raise ValueError(f"Unsupported safetensors format: {safetensor_path}")
+        
     write_combined_tensor(target_tensor_map, output_path, gguf_loader)
     
     return

setup.py

@@ -673,10 +673,29 @@ if not torch.xpu.is_available() and not KTRANSFORMERS_BUILD_NPU:
         ext_modules.append(
             CMakeExtension("balance_serve", os.fspath(Path("").resolve()/ "csrc"/ "balance_serve"))
         )
+
+    setup(
+        name=VersionInfo.PACKAGE_NAME,
+        version=VersionInfo().get_package_version(),
+        install_requires=triton_dep,
+        cmdclass={"bdist_wheel":BuildWheelsCommand ,"build_ext": CMakeBuild},
+        ext_modules=ext_modules
+    )
+
+
+
 elif torch.xpu.is_available():
     ext_modules = [
         CMakeExtension("cpuinfer_ext", os.fspath(Path("").resolve() / "csrc" / "ktransformers_ext")),
     ]
+    setup(
+        name=VersionInfo.PACKAGE_NAME,
+        version=VersionInfo().get_package_version(),
+        install_requires=triton_dep,
+        cmdclass={"bdist_wheel":BuildWheelsCommand ,"build_ext": CMakeBuild},
+        ext_modules=ext_modules
+    )
+
 elif KTRANSFORMERS_BUILD_NPU:
     ext_modules = [
         CMakeExtension("cpuinfer_ext", os.fspath(Path("").resolve() / "csrc" / "ktransformers_ext")),
@@ -687,11 +706,10 @@ elif KTRANSFORMERS_BUILD_NPU:
             CMakeExtension("balance_serve", os.fspath(Path("").resolve()/ "csrc"/ "balance_serve"))
         )
 
+    setup(
+        name=VersionInfo.PACKAGE_NAME,
+        version=VersionInfo().get_package_version(),
+        cmdclass={"bdist_wheel":BuildWheelsCommand ,"build_ext": CMakeBuild},
+        ext_modules=ext_modules
+    )
 
-setup(
-    name=VersionInfo.PACKAGE_NAME,
-    version=VersionInfo().get_package_version(),
-    install_requires=triton_dep,
-    cmdclass={"bdist_wheel":BuildWheelsCommand ,"build_ext": CMakeBuild},
-    ext_modules=ext_modules
-)

third_party/llamafile/iqk_mul_mat_arm80.cpp

@@ -0,0 +1,10 @@
+// Adapted from
+// https://github.com/Mozilla-Ocho/llamafile/blob/0.8.8/llamafile/iqk_mul_mat_arm80.cpp
+// Copyright 2024 Iwan Kawrakow.
+// Copyright(c) 2024 by KVCache.AI, All Rights Reserved.
+
+#ifdef __aarch64__
+#define iqk_mul_mat iqk_mul_mat_arm80
+#define iqk_mul_mat_moe iqk_mul_mat_moe_arm80
+#include "iqk_mul_mat.inc"
+#endif // __aarch64__

third_party/llamafile/sgemm.cpp

@@ -1,204 +1,7 @@
-// Adapted from
-// https://github.com/Mozilla-Ocho/llamafile/blob/0.8.8/llamafile/sgemm.cpp
-// Copyrigth 2024 Mozilla Foundation.
-// Copyright(c) 2024 by KVCache.AI, All Rights Reserved.
-
-// -*- mode:c++;indent-tabs-mode:nil;c-basic-offset:4;coding:utf-8 -*-
-// vi: set et ft=cpp ts=4 sts=4 sw=4 fenc=utf-8 :vi
-//
-// Copyright 2024 Mozilla Foundation
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#include "sgemm.h"
-// #include <cosmo.h>
-// #include <cpuid.h>
-// #include <libc/sysv/consts/hwcap.h>
-#include <stdio.h>
-// #include <sys/auxv.h>
-#include <cassert>
-// #include "llamafile.h"
-
-static const struct GemmFuncs {
-    bool (*sgemm)(long, long, long, const void*, long, const void*, long, void*, long, int, int, int, int, int, int, int);
-    bool (*mixmul)(const struct ggml_compute_params*, const struct ggml_tensor*, const struct ggml_tensor*, const struct ggml_tensor*, struct ggml_tensor*);
-    bool (*iqk_mixmul)(long, long, long, int, int, const void*, const void*, float*, long, long, const void*, int, int);
-    // typeof(llamafile_sgemm)* sgemm;
-    // typeof(llamafile_mixmul)* mixmul;
-    // typeof(llamafile_mixmul_iqk)* iqk_mixmul = iqk_mul_mat_moe_unsupported;
-    GemmFuncs() {
-#if defined(__x86_64__) || defined(_M_X64)
-        // if (X86_HAVE(AVX)) {
-        //     if (X86_HAVE(FMA)) {
-        //         if (X86_HAVE(AVX2)) {
-        //             if (X86_HAVE(AVX512F)) {
-        //                 if (X86_HAVE(AVX512VL) &&     //
-        //                     X86_HAVE(AVX512BW) &&     //
-        //                     X86_HAVE(AVX512DQ) &&     //
-        //                     X86_HAVE(AVX512_VNNI) &&  //
-        //                     X86_HAVE(AVX512_BF16)) {
-        //                     // AMD Zen4+ (2023-)
-        //                     sgemm = llamafile_sgemm_amd_zen4;
-        //                     mixmul = llamafile_mixmul_amd_zen4;
-        //                     iqk_mixmul = iqk_mul_mat_moe_zen4;
-        //                 } else {
-        //                     // Intel Xeon Skylake+ (2015-)
-        //                     sgemm = llamafile_sgemm_amd_avx512f;
-        //                     mixmul = llamafile_mixmul_amd_avx512f;
-        //                     iqk_mixmul = iqk_mul_mat_moe;
-        //                 }
-        //             } else if (X86_HAVE(AVXVNNI)) {
-        //                 // Intel Alderlake (2021-)
-        //                 sgemm = llamafile_sgemm_amd_avxvnni;
-        //                 mixmul = llamafile_mixmul_amd_avxvnni;
-        //                 iqk_mixmul = iqk_mul_mat_moe;
-        //             } else {
-        //                 // Intel Haswell/Broadwell/Skylake (2013-2020)
-        //                 // AMD Excavator (2015-2022)
-        //                 sgemm = llamafile_sgemm_amd_avx2;
-        //                 mixmul = llamafile_mixmul_amd_avx2;
-        //                 if (X86_HAVE(F16C))
-        //                     iqk_mixmul = iqk_mul_mat_moe;
-        //             }
-        //         } else {
-        //             // AMD Piledriver (2011-2014)
-        //             sgemm = llamafile_sgemm_amd_fma;
-        //             mixmul = llamafile_mixmul_amd_fma;
-        //             if (X86_HAVE(F16C))
-        //                 iqk_mixmul = iqk_mul_mat_moe;
-        //         }
-        //     } else {
-        //         // Intel Sandybridge/Ivybridge (2010-2012)
-        //         // AMD Bulldozer (2011)
-        //         sgemm = llamafile_sgemm_amd_avx;
-        //         mixmul = llamafile_mixmul_amd_avx;
-        //     }
-        // } else {
-        //     // AMD K8/Barcelona (2003-2010)
-        //     // Intel Core/Nehalem (2006-2009)
-        //     sgemm = llamafile_sgemm_unsupported;
-        //     mixmul = llamafile_mixmul_unsupported;
-        // }
-
-#if defined(__AVX__)
-#if defined(__FMA__) || (defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__)))
-#if defined(__AVX2__)
-#if defined(__AVX512F__)
-#if defined(__AVX512VL__) && defined(__AVX512BW__) && defined(__AVX512DQ__) && defined(__AVX512VNNI__) && defined(__AVX512BF16__)
-        // AMD Zen4+ (2023-)
-        sgemm = llamafile_sgemm_amd_zen4;
-        mixmul = llamafile_mixmul_amd_zen4;
-        iqk_mixmul = iqk_mul_mat_moe_zen4;
+#if defined(KTRANSFORMERS_USE_NPU) && KTRANSFORMERS_USE_NPU
+    // 使用 x86 版本
+    #include "sgemm_arm.cpp"
 #else
-        // Intel Xeon Skylake+ (2015-)
-        sgemm = llamafile_sgemm_amd_avx512f;
-        mixmul = llamafile_mixmul_amd_avx512f;
-        iqk_mixmul = iqk_mul_mat_moe;
-#endif
-#elif defined(__AVXVNNI__)
-        // Intel Alderlake (2021-)
-        sgemm = llamafile_sgemm_amd_avxvnni;
-        mixmul = llamafile_mixmul_amd_avxvnni;
-        iqk_mixmul = iqk_mul_mat_moe;
-#else
-        // Intel Haswell/Broadwell/Skylake (2013-2020)
-        // AMD Excavator (2015-2022)
-        sgemm = llamafile_sgemm_amd_avx2;
-        mixmul = llamafile_mixmul_amd_avx2;
-#if defined(__F16C__)
-        iqk_mixmul = iqk_mul_mat_moe;
-#endif
-#endif
-#else
-        // AMD Piledriver (2011-2014)
-        sgemm = llamafile_sgemm_amd_fma;
-        mixmul = llamafile_mixmul_amd_fma;
-#if defined(__F16C__)
-        iqk_mixmul = iqk_mul_mat_moe;
-#endif
-#endif
-#else
-        // Intel Sandybridge/Ivybridge (2010-2012)
-        // AMD Bulldozer (2011)
-        sgemm = llamafile_sgemm_amd_avx;
-        mixmul = llamafile_mixmul_amd_avx;
-#endif
-#else
-        // AMD K8/Barcelona (2003-2010)
-        // Intel Core/Nehalem (2006-2009)
-        sgemm = llamafile_sgemm_unsupported;
-        mixmul = llamafile_mixmul_unsupported;
-#endif
-
-#elif defined(__aarch64__)
-        long hwcap = getauxval(AT_HWCAP);
-        if ((hwcap & HWCAP_FPHP) &&     // fp16 scalar isa (ID_AA64PFR0_EL1.FP == 1)
-            (hwcap & HWCAP_ASIMDHP) &&  // fp16 vector isa (ID_AA64PFR0_EL1.AdvSIMD == 1)
-            (hwcap & HWCAP_ASIMDDP)) {  // dotprod isa (ID_AA64ISAR0_EL1.DP == 1)
-            // e.g. Apple M1, Raspberry Pi 5
-            sgemm = llamafile_sgemm_arm82;
-            mixmul = llamafile_mixmul_arm82;
-            iqk_mixmul = iqk_mul_mat_moe_arm82;
-        } else {
-            // ARM64 baseline ISA
-            sgemm = llamafile_sgemm_arm80;
-            mixmul = llamafile_mixmul_arm80;
-        }
-#else
-        sgemm = llamafile_sgemm_unsupported;
-        mixmul = llamafile_mixmul_unsupported;
-#endif
-    }
-} funcs;
-
-/**
- * Performs optimized matrix multiplication on CPU.
- *
- * This subroutine may compute C = Aᵀ * B with column major ordering.
- * Despite its name, this isn't a generalized implementation. Work is
- * only performed when a handwritten kernel is written and available.
- * Otherwise the caller should fall back to a general matmul routine.
- *
- * @param m is rows in `A` and `C`
- * @param n is cols in `B` and `C`
- * @param k is cols in `A` and rows in `B`
- * @param A is first input matrix (always transposed)
- * @param lda is row stride of `A`
- * @param B is second input matrix (never transposed)
- * @param ldb is row stride of `B`
- * @param C is input/output array of output matrices
- * @param ldc is row stride of `C`
- * @param ith is thread id (must be less than `nth`)
- * @param nth is number of threads (must be greater than zero)
- * @param task is GGML task type
- * @param Atype is GGML data type of `A`
- * @param Btype is GGML data type of `B`
- * @param Ctype is GGML data type of `C`
- * @param precision may be used to control the internal compute type
- * @return true if this function was able to service the matmul request
- */
-bool llamafile_sgemm(long m, long n, long k, const void* A, long lda, const void* B, long ldb, void* C, long ldc, int ith, int nth, int task, int Atype, int Btype, int Ctype, int precision) {
-    return funcs.sgemm(m, n, k, A, lda, B, ldb, C, ldc, ith, nth, task, Atype, Btype, Ctype,
-                       precision);
-}
-
-/**
- * Performs "mixture of experts" tensor multiplication on CPU.
- */
-bool llamafile_mixmul(const ggml_compute_params* params, const ggml_tensor* weights, const ggml_tensor* thought, const ggml_tensor* plan, ggml_tensor* result) {
-    return funcs.mixmul(params, weights, thought, plan, result);
-}
-
-bool llamafile_mixmul_iqk(long Nx, long Ny, long ne00, int ne11, int typeA, const void* A, const void* B, float* C, long nb1, long nb2, const void* vrow_mapping, int ith, int nth) {
-    return funcs.iqk_mixmul(Nx, Ny, ne00, ne11, typeA, A, B, C, nb1, nb2, vrow_mapping, ith, nth);
-}
+    // 使用 ARM 版本
+    #include "sgemm_x86.cpp"
+#endif

third_party/llamafile/sgemm_arm.cpp

@@ -0,0 +1,204 @@
+// Adapted from
+// https://github.com/Mozilla-Ocho/llamafile/blob/0.8.8/llamafile/sgemm.cpp
+// Copyrigth 2024 Mozilla Foundation.
+// Copyright(c) 2024 by KVCache.AI, All Rights Reserved.
+
+// -*- mode:c++;indent-tabs-mode:nil;c-basic-offset:4;coding:utf-8 -*-
+// vi: set et ft=cpp ts=4 sts=4 sw=4 fenc=utf-8 :vi
+//
+// Copyright 2024 Mozilla Foundation
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "sgemm.h"
+// #include <cosmo.h>
+// #include <cpuid.h>
+// #include <libc/sysv/consts/hwcap.h>
+#include <stdio.h>
+// #include <sys/auxv.h>
+#include <cassert>
+// #include "llamafile.h"
+
+static const struct GemmFuncs {
+    bool (*sgemm)(long, long, long, const void*, long, const void*, long, void*, long, int, int, int, int, int, int, int);
+    bool (*mixmul)(const struct ggml_compute_params*, const struct ggml_tensor*, const struct ggml_tensor*, const struct ggml_tensor*, struct ggml_tensor*);
+    bool (*iqk_mixmul)(long, long, long, int, int, const void*, const void*, float*, long, long, const void*, int, int);
+    // typeof(llamafile_sgemm)* sgemm;
+    // typeof(llamafile_mixmul)* mixmul;
+    // typeof(llamafile_mixmul_iqk)* iqk_mixmul = iqk_mul_mat_moe_unsupported;
+    GemmFuncs() {
+#if defined(__x86_64__) || defined(_M_X64)
+        // if (X86_HAVE(AVX)) {
+        //     if (X86_HAVE(FMA)) {
+        //         if (X86_HAVE(AVX2)) {
+        //             if (X86_HAVE(AVX512F)) {
+        //                 if (X86_HAVE(AVX512VL) &&     //
+        //                     X86_HAVE(AVX512BW) &&     //
+        //                     X86_HAVE(AVX512DQ) &&     //
+        //                     X86_HAVE(AVX512_VNNI) &&  //
+        //                     X86_HAVE(AVX512_BF16)) {
+        //                     // AMD Zen4+ (2023-)
+        //                     sgemm = llamafile_sgemm_amd_zen4;
+        //                     mixmul = llamafile_mixmul_amd_zen4;
+        //                     iqk_mixmul = iqk_mul_mat_moe_zen4;
+        //                 } else {
+        //                     // Intel Xeon Skylake+ (2015-)
+        //                     sgemm = llamafile_sgemm_amd_avx512f;
+        //                     mixmul = llamafile_mixmul_amd_avx512f;
+        //                     iqk_mixmul = iqk_mul_mat_moe;
+        //                 }
+        //             } else if (X86_HAVE(AVXVNNI)) {
+        //                 // Intel Alderlake (2021-)
+        //                 sgemm = llamafile_sgemm_amd_avxvnni;
+        //                 mixmul = llamafile_mixmul_amd_avxvnni;
+        //                 iqk_mixmul = iqk_mul_mat_moe;
+        //             } else {
+        //                 // Intel Haswell/Broadwell/Skylake (2013-2020)
+        //                 // AMD Excavator (2015-2022)
+        //                 sgemm = llamafile_sgemm_amd_avx2;
+        //                 mixmul = llamafile_mixmul_amd_avx2;
+        //                 if (X86_HAVE(F16C))
+        //                     iqk_mixmul = iqk_mul_mat_moe;
+        //             }
+        //         } else {
+        //             // AMD Piledriver (2011-2014)
+        //             sgemm = llamafile_sgemm_amd_fma;
+        //             mixmul = llamafile_mixmul_amd_fma;
+        //             if (X86_HAVE(F16C))
+        //                 iqk_mixmul = iqk_mul_mat_moe;
+        //         }
+        //     } else {
+        //         // Intel Sandybridge/Ivybridge (2010-2012)
+        //         // AMD Bulldozer (2011)
+        //         sgemm = llamafile_sgemm_amd_avx;
+        //         mixmul = llamafile_mixmul_amd_avx;
+        //     }
+        // } else {
+        //     // AMD K8/Barcelona (2003-2010)
+        //     // Intel Core/Nehalem (2006-2009)
+        //     sgemm = llamafile_sgemm_unsupported;
+        //     mixmul = llamafile_mixmul_unsupported;
+        // }
+
+#if defined(__AVX__)
+#if defined(__FMA__) || (defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__)))
+#if defined(__AVX2__)
+#if defined(__AVX512F__)
+#if defined(__AVX512VL__) && defined(__AVX512BW__) && defined(__AVX512DQ__) && defined(__AVX512VNNI__) && defined(__AVX512BF16__)
+        // AMD Zen4+ (2023-)
+        sgemm = llamafile_sgemm_amd_zen4;
+        mixmul = llamafile_mixmul_amd_zen4;
+        iqk_mixmul = iqk_mul_mat_moe_zen4;
+#else
+        // Intel Xeon Skylake+ (2015-)
+        sgemm = llamafile_sgemm_amd_avx512f;
+        mixmul = llamafile_mixmul_amd_avx512f;
+        iqk_mixmul = iqk_mul_mat_moe;
+#endif
+#elif defined(__AVXVNNI__)
+        // Intel Alderlake (2021-)
+        sgemm = llamafile_sgemm_amd_avxvnni;
+        mixmul = llamafile_mixmul_amd_avxvnni;
+        iqk_mixmul = iqk_mul_mat_moe;
+#else
+        // Intel Haswell/Broadwell/Skylake (2013-2020)
+        // AMD Excavator (2015-2022)
+        sgemm = llamafile_sgemm_amd_avx2;
+        mixmul = llamafile_mixmul_amd_avx2;
+#if defined(__F16C__)
+        iqk_mixmul = iqk_mul_mat_moe;
+#endif
+#endif
+#else
+        // AMD Piledriver (2011-2014)
+        sgemm = llamafile_sgemm_amd_fma;
+        mixmul = llamafile_mixmul_amd_fma;
+#if defined(__F16C__)
+        iqk_mixmul = iqk_mul_mat_moe;
+#endif
+#endif
+#else
+        // Intel Sandybridge/Ivybridge (2010-2012)
+        // AMD Bulldozer (2011)
+        sgemm = llamafile_sgemm_amd_avx;
+        mixmul = llamafile_mixmul_amd_avx;
+#endif
+#else
+        // AMD K8/Barcelona (2003-2010)
+        // Intel Core/Nehalem (2006-2009)
+        sgemm = llamafile_sgemm_unsupported;
+        mixmul = llamafile_mixmul_unsupported;
+#endif
+
+#elif defined(__aarch64__)
+//        long hwcap = getauxval(AT_HWCAP);
+//        if ((hwcap & HWCAP_FPHP) &&     // fp16 scalar isa (ID_AA64PFR0_EL1.FP == 1)
+//            (hwcap & HWCAP_ASIMDHP) &&  // fp16 vector isa (ID_AA64PFR0_EL1.AdvSIMD == 1)
+//            (hwcap & HWCAP_ASIMDDP)) {  // dotprod isa (ID_AA64ISAR0_EL1.DP == 1)
+//            // e.g. Apple M1, Raspberry Pi 5
+//            sgemm = llamafile_sgemm_arm82;
+//            mixmul = llamafile_mixmul_arm82;
+//            iqk_mixmul = iqk_mul_mat_moe_arm82;
+//        } else {
+            // ARM64 baseline ISA
+            sgemm = llamafile_sgemm_arm80;
+            mixmul = llamafile_mixmul_arm80;
+//        }
+#else
+        sgemm = llamafile_sgemm_unsupported;
+        mixmul = llamafile_mixmul_unsupported;
+#endif
+    }
+} funcs;
+
+/**
+ * Performs optimized matrix multiplication on CPU.
+ *
+ * This subroutine may compute C = Aᵀ * B with column major ordering.
+ * Despite its name, this isn't a generalized implementation. Work is
+ * only performed when a handwritten kernel is written and available.
+ * Otherwise the caller should fall back to a general matmul routine.
+ *
+ * @param m is rows in `A` and `C`
+ * @param n is cols in `B` and `C`
+ * @param k is cols in `A` and rows in `B`
+ * @param A is first input matrix (always transposed)
+ * @param lda is row stride of `A`
+ * @param B is second input matrix (never transposed)
+ * @param ldb is row stride of `B`
+ * @param C is input/output array of output matrices
+ * @param ldc is row stride of `C`
+ * @param ith is thread id (must be less than `nth`)
+ * @param nth is number of threads (must be greater than zero)
+ * @param task is GGML task type
+ * @param Atype is GGML data type of `A`
+ * @param Btype is GGML data type of `B`
+ * @param Ctype is GGML data type of `C`
+ * @param precision may be used to control the internal compute type
+ * @return true if this function was able to service the matmul request
+ */
+bool llamafile_sgemm(long m, long n, long k, const void* A, long lda, const void* B, long ldb, void* C, long ldc, int ith, int nth, int task, int Atype, int Btype, int Ctype, int precision) {
+    return funcs.sgemm(m, n, k, A, lda, B, ldb, C, ldc, ith, nth, task, Atype, Btype, Ctype,
+                       precision);
+}
+
+/**
+ * Performs "mixture of experts" tensor multiplication on CPU.
+ */
+bool llamafile_mixmul(const ggml_compute_params* params, const ggml_tensor* weights, const ggml_tensor* thought, const ggml_tensor* plan, ggml_tensor* result) {
+    return funcs.mixmul(params, weights, thought, plan, result);
+}
+
+bool llamafile_mixmul_iqk(long Nx, long Ny, long ne00, int ne11, int typeA, const void* A, const void* B, float* C, long nb1, long nb2, const void* vrow_mapping, int ith, int nth) {
+    return funcs.iqk_mixmul(Nx, Ny, ne00, ne11, typeA, A, B, C, nb1, nb2, vrow_mapping, ith, nth);
+}

third_party/llamafile/sgemm_x86.cpp

@@ -0,0 +1,204 @@
+// Adapted from
+// https://github.com/Mozilla-Ocho/llamafile/blob/0.8.8/llamafile/sgemm.cpp
+// Copyrigth 2024 Mozilla Foundation.
+// Copyright(c) 2024 by KVCache.AI, All Rights Reserved.
+
+// -*- mode:c++;indent-tabs-mode:nil;c-basic-offset:4;coding:utf-8 -*-
+// vi: set et ft=cpp ts=4 sts=4 sw=4 fenc=utf-8 :vi
+//
+// Copyright 2024 Mozilla Foundation
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "sgemm.h"
+// #include <cosmo.h>
+// #include <cpuid.h>
+// #include <libc/sysv/consts/hwcap.h>
+#include <stdio.h>
+// #include <sys/auxv.h>
+#include <cassert>
+// #include "llamafile.h"
+
+static const struct GemmFuncs {
+    bool (*sgemm)(long, long, long, const void*, long, const void*, long, void*, long, int, int, int, int, int, int, int);
+    bool (*mixmul)(const struct ggml_compute_params*, const struct ggml_tensor*, const struct ggml_tensor*, const struct ggml_tensor*, struct ggml_tensor*);
+    bool (*iqk_mixmul)(long, long, long, int, int, const void*, const void*, float*, long, long, const void*, int, int);
+    // typeof(llamafile_sgemm)* sgemm;
+    // typeof(llamafile_mixmul)* mixmul;
+    // typeof(llamafile_mixmul_iqk)* iqk_mixmul = iqk_mul_mat_moe_unsupported;
+    GemmFuncs() {
+#if defined(__x86_64__) || defined(_M_X64)
+        // if (X86_HAVE(AVX)) {
+        //     if (X86_HAVE(FMA)) {
+        //         if (X86_HAVE(AVX2)) {
+        //             if (X86_HAVE(AVX512F)) {
+        //                 if (X86_HAVE(AVX512VL) &&     //
+        //                     X86_HAVE(AVX512BW) &&     //
+        //                     X86_HAVE(AVX512DQ) &&     //
+        //                     X86_HAVE(AVX512_VNNI) &&  //
+        //                     X86_HAVE(AVX512_BF16)) {
+        //                     // AMD Zen4+ (2023-)
+        //                     sgemm = llamafile_sgemm_amd_zen4;
+        //                     mixmul = llamafile_mixmul_amd_zen4;
+        //                     iqk_mixmul = iqk_mul_mat_moe_zen4;
+        //                 } else {
+        //                     // Intel Xeon Skylake+ (2015-)
+        //                     sgemm = llamafile_sgemm_amd_avx512f;
+        //                     mixmul = llamafile_mixmul_amd_avx512f;
+        //                     iqk_mixmul = iqk_mul_mat_moe;
+        //                 }
+        //             } else if (X86_HAVE(AVXVNNI)) {
+        //                 // Intel Alderlake (2021-)
+        //                 sgemm = llamafile_sgemm_amd_avxvnni;
+        //                 mixmul = llamafile_mixmul_amd_avxvnni;
+        //                 iqk_mixmul = iqk_mul_mat_moe;
+        //             } else {
+        //                 // Intel Haswell/Broadwell/Skylake (2013-2020)
+        //                 // AMD Excavator (2015-2022)
+        //                 sgemm = llamafile_sgemm_amd_avx2;
+        //                 mixmul = llamafile_mixmul_amd_avx2;
+        //                 if (X86_HAVE(F16C))
+        //                     iqk_mixmul = iqk_mul_mat_moe;
+        //             }
+        //         } else {
+        //             // AMD Piledriver (2011-2014)
+        //             sgemm = llamafile_sgemm_amd_fma;
+        //             mixmul = llamafile_mixmul_amd_fma;
+        //             if (X86_HAVE(F16C))
+        //                 iqk_mixmul = iqk_mul_mat_moe;
+        //         }
+        //     } else {
+        //         // Intel Sandybridge/Ivybridge (2010-2012)
+        //         // AMD Bulldozer (2011)
+        //         sgemm = llamafile_sgemm_amd_avx;
+        //         mixmul = llamafile_mixmul_amd_avx;
+        //     }
+        // } else {
+        //     // AMD K8/Barcelona (2003-2010)
+        //     // Intel Core/Nehalem (2006-2009)
+        //     sgemm = llamafile_sgemm_unsupported;
+        //     mixmul = llamafile_mixmul_unsupported;
+        // }
+
+#if defined(__AVX__)
+#if defined(__FMA__) || (defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__)))
+#if defined(__AVX2__)
+#if defined(__AVX512F__)
+#if defined(__AVX512VL__) && defined(__AVX512BW__) && defined(__AVX512DQ__) && defined(__AVX512VNNI__) && defined(__AVX512BF16__)
+        // AMD Zen4+ (2023-)
+        sgemm = llamafile_sgemm_amd_zen4;
+        mixmul = llamafile_mixmul_amd_zen4;
+        iqk_mixmul = iqk_mul_mat_moe_zen4;
+#else
+        // Intel Xeon Skylake+ (2015-)
+        sgemm = llamafile_sgemm_amd_avx512f;
+        mixmul = llamafile_mixmul_amd_avx512f;
+        iqk_mixmul = iqk_mul_mat_moe;
+#endif
+#elif defined(__AVXVNNI__)
+        // Intel Alderlake (2021-)
+        sgemm = llamafile_sgemm_amd_avxvnni;
+        mixmul = llamafile_mixmul_amd_avxvnni;
+        iqk_mixmul = iqk_mul_mat_moe;
+#else
+        // Intel Haswell/Broadwell/Skylake (2013-2020)
+        // AMD Excavator (2015-2022)
+        sgemm = llamafile_sgemm_amd_avx2;
+        mixmul = llamafile_mixmul_amd_avx2;
+#if defined(__F16C__)
+        iqk_mixmul = iqk_mul_mat_moe;
+#endif
+#endif
+#else
+        // AMD Piledriver (2011-2014)
+        sgemm = llamafile_sgemm_amd_fma;
+        mixmul = llamafile_mixmul_amd_fma;
+#if defined(__F16C__)
+        iqk_mixmul = iqk_mul_mat_moe;
+#endif
+#endif
+#else
+        // Intel Sandybridge/Ivybridge (2010-2012)
+        // AMD Bulldozer (2011)
+        sgemm = llamafile_sgemm_amd_avx;
+        mixmul = llamafile_mixmul_amd_avx;
+#endif
+#else
+        // AMD K8/Barcelona (2003-2010)
+        // Intel Core/Nehalem (2006-2009)
+        sgemm = llamafile_sgemm_unsupported;
+        mixmul = llamafile_mixmul_unsupported;
+#endif
+
+#elif defined(__aarch64__)
+        long hwcap = getauxval(AT_HWCAP);
+        if ((hwcap & HWCAP_FPHP) &&     // fp16 scalar isa (ID_AA64PFR0_EL1.FP == 1)
+            (hwcap & HWCAP_ASIMDHP) &&  // fp16 vector isa (ID_AA64PFR0_EL1.AdvSIMD == 1)
+            (hwcap & HWCAP_ASIMDDP)) {  // dotprod isa (ID_AA64ISAR0_EL1.DP == 1)
+            // e.g. Apple M1, Raspberry Pi 5
+            sgemm = llamafile_sgemm_arm82;
+            mixmul = llamafile_mixmul_arm82;
+            iqk_mixmul = iqk_mul_mat_moe_arm82;
+        } else {
+            // ARM64 baseline ISA
+            sgemm = llamafile_sgemm_arm80;
+            mixmul = llamafile_mixmul_arm80;
+        }
+#else
+        sgemm = llamafile_sgemm_unsupported;
+        mixmul = llamafile_mixmul_unsupported;
+#endif
+    }
+} funcs;
+
+/**
+ * Performs optimized matrix multiplication on CPU.
+ *
+ * This subroutine may compute C = Aᵀ * B with column major ordering.
+ * Despite its name, this isn't a generalized implementation. Work is
+ * only performed when a handwritten kernel is written and available.
+ * Otherwise the caller should fall back to a general matmul routine.
+ *
+ * @param m is rows in `A` and `C`
+ * @param n is cols in `B` and `C`
+ * @param k is cols in `A` and rows in `B`
+ * @param A is first input matrix (always transposed)
+ * @param lda is row stride of `A`
+ * @param B is second input matrix (never transposed)
+ * @param ldb is row stride of `B`
+ * @param C is input/output array of output matrices
+ * @param ldc is row stride of `C`
+ * @param ith is thread id (must be less than `nth`)
+ * @param nth is number of threads (must be greater than zero)
+ * @param task is GGML task type
+ * @param Atype is GGML data type of `A`
+ * @param Btype is GGML data type of `B`
+ * @param Ctype is GGML data type of `C`
+ * @param precision may be used to control the internal compute type
+ * @return true if this function was able to service the matmul request
+ */
+bool llamafile_sgemm(long m, long n, long k, const void* A, long lda, const void* B, long ldb, void* C, long ldc, int ith, int nth, int task, int Atype, int Btype, int Ctype, int precision) {
+    return funcs.sgemm(m, n, k, A, lda, B, ldb, C, ldc, ith, nth, task, Atype, Btype, Ctype,
+                       precision);
+}
+
+/**
+ * Performs "mixture of experts" tensor multiplication on CPU.
+ */
+bool llamafile_mixmul(const ggml_compute_params* params, const ggml_tensor* weights, const ggml_tensor* thought, const ggml_tensor* plan, ggml_tensor* result) {
+    return funcs.mixmul(params, weights, thought, plan, result);
+}
+
+bool llamafile_mixmul_iqk(long Nx, long Ny, long ne00, int ne11, int typeA, const void* A, const void* B, float* C, long nb1, long nb2, const void* vrow_mapping, int ith, int nth) {
+    return funcs.iqk_mixmul(Nx, Ny, ne00, ne11, typeA, A, B, C, nb1, nb2, vrow_mapping, ith, nth);
+}

third_party/llamafile/tinyblas_cpu_mixmul_arm80.cpp

@@ -5,6 +5,7 @@
 
 #ifdef __aarch64__
 #define llamafile_mixmul llamafile_mixmul_arm80
+#define iqk_mul_mat iqk_mul_mat_arm80
 #include "tinyblas_cpu_mixmul.inc"
 
 /**

third_party/llamafile/tinyblas_cpu_sgemm.inc

@@ -1,361 +1,7 @@
-// Adapted from
-// https://github.com/Mozilla-Ocho/llamafile/blob/0.8.8/llamafile/tinyblas_cpu_sgemm.inc
-// Copyrigth 2024 Mozilla Foundation.
-// Copyright(c) 2024 by KVCache.AI, All Rights Reserved.
-
-// -*- mode:c++;indent-tabs-mode:nil;c-basic-offset:4;coding:utf-8 -*-
-// vi: set et ft=cpp ts=4 sts=4 sw=4 fenc=utf-8 :vi
-//
-// Copyright 2024 Mozilla Foundation
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#include "tinyblas_cpu.h"
-
-//
-//
-//                                ██████╗ ██╗   █████╗ ██████╗
-//         ██████╗██╗██╗ ██╗██═██╗██╔══██╗██║  ██╔══██╗██╔═══╝
-//         ╚═██╔═╝██║███▄██║██ ██║██████╔╝██║  ███████║██████╗
-//           ██║  ██║██▀███║╚███╔╝██╔══██╗██║  ██╔══██║╔═══██║
-//           ██║  ██║██║ ██║ ███║ ██████╔╝████╗██║  ██║██████║
-//           ╚═╝  ╚═╝╚═╝ ╚═╝ ╚══╝ ╚═════╝ ╚═══╝╚═╝  ╚═╝╚═════╝
-//
-//                   BASIC LINEAR ALGEBRA SUBPROGRAMS
-//
-//
-// This file implements multithreaded CPU matrix multiplication for the
-// common contiguous use case C = Aᵀ * B. These kernels are designed to
-// have excellent performance[1] for matrices that fit in the CPU cache
-// without imposing any overhead such as cache filling or malloc calls.
-//
-// This implementation does not guarantee any upper bound with rounding
-// errors, which grow along with k. Our goal's to maximally exploit the
-// hardware for performance, and then use whatever resources remain for
-// improving numerical accuracy.
-//
-// [1] J. Tunney, ‘LLaMA Now Goes Faster on CPUs’, Mar. 2024. [Online].
-//     Available: https://justine.lol/matmul/. [Accessed: 29-Mar-2024].
-
-namespace {
-
-template <typename TC>
-bool llamafile_sgemm_impl(long m, long n, long k, const void* A, long lda, const void* B, long ldb, TC* C, long ldc, int ith, int nth, int task, int Atype, int Btype, int Ctype, int precision) {
-    switch (Atype) {
-        case GGML_TYPE_F32: {
-            if (Btype != GGML_TYPE_F32)
-                return NOT_SUPPORTED;
-#if defined(__AVX512F__)
-            if (k % 16)
-                return NOT_SUPPORTED;
-            tinyBLAS<0, 16, __m512, __m512, float, float, TC> tb{
-                k, (const float*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
-            tb.matmul(m, n, task);
-            return true;
-#elif defined(__AVX__) || defined(__AVX2__)
-            if (k % 8)
-                return NOT_SUPPORTED;
-            tinyBLAS<0, 8, __m256, __m256, float, float, TC> tb{
-                k, (const float*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
-            tb.matmul(m, n, task);
-            return true;
-#elif defined(__ARM_NEON)
-            if (k % 4)
-                return NOT_SUPPORTED;
-            tinyBLAS<0, 4, float32x4_t, float32x4_t, float, float, TC> tb{
-                k, (const float*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
-            tb.matmul(m, n, task);
-            return true;
+#if defined(KTRANSFORMERS_USE_NPU) && KTRANSFORMERS_USE_NPU
+    // 使用 x86 版本
+    #include "tinyblas_cpu_sgemm_arm.inc"
 #else
-            return NOT_SUPPORTED;
-#endif
-        }
-
-        case GGML_TYPE_BF16: {
-#if defined(__AVX512BF16__)
-            if (k % 32)
-                return NOT_SUPPORTED;
-            if (Btype == GGML_TYPE_F32 && n < 2) {
-                tinyBLAS<0, 16, __m512, __m512, ggml_bf16_t, float, TC> tb{
-                    k, (const ggml_bf16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
-                tb.matmul(m, n, task);
-                return true;
-            }
-            if (Btype == GGML_TYPE_F32)
-                return WANT_QUANTIZATION;
-            if (Btype != GGML_TYPE_BF16)
-                return NOT_SUPPORTED;
-            if (!FLAG_precise) {
-                tinyBLAS<0, 32, __m512, __m512bh, ggml_bf16_t, ggml_bf16_t, TC> tb{
-                    k, (const ggml_bf16_t*)A, lda, (const ggml_bf16_t*)B, ldb, C, ldc, ith, nth};
-                tb.matmul(m, n, task);
-                return true;
-            } else {
-                tinyBLAS<0, 16, __m512, __m512, ggml_bf16_t, ggml_bf16_t, TC> tb{
-                    k, (const ggml_bf16_t*)A, lda, (const ggml_bf16_t*)B, ldb, C, ldc, ith, nth};
-                tb.matmul(m, n, task);
-                return true;
-            }
-#elif defined(__AVX512F__)
-            if (k % 16)
-                return NOT_SUPPORTED;
-            tinyBLAS<0, 16, __m512, __m512, ggml_bf16_t, float, TC> tb{
-                k, (const ggml_bf16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
-            tb.matmul(m, n, task);
-            return true;
-#elif defined(__AVX2__)
-            if (k % 8)
-                return NOT_SUPPORTED;
-            if (Btype != GGML_TYPE_F32)
-                return NOT_SUPPORTED;
-            tinyBLAS<0, 8, __m256, __m256, ggml_bf16_t, float, TC> tb{
-                k, (const ggml_bf16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
-            tb.matmul(m, n, task);
-            return true;
-#elif defined(__ARM_NEON) && !defined(_MSC_VER)
-            if (k % 4)
-                return NOT_SUPPORTED;
-            if (Btype != GGML_TYPE_F32)
-                return NOT_SUPPORTED;
-            tinyBLAS<0, 4, float32x4_t, float32x4_t, ggml_bf16_t, float, TC> tb{
-                k, (const ggml_bf16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
-            tb.matmul(m, n, task);
-            return true;
-#else
-            return NOT_SUPPORTED;
-#endif
-        }
-
-        case GGML_TYPE_F16: {
-#if defined(__AVX512F__)
-            if (k % 16)
-                return NOT_SUPPORTED;
-            if (Btype == GGML_TYPE_F32 && n < 2) {
-                tinyBLAS<0, 16, __m512, __m512, ggml_fp16_t, float, TC> tb{
-                    k, (const ggml_fp16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
-                tb.matmul(m, n, task);
-                return true;
-            }
-            if (Btype == GGML_TYPE_F32)
-                return WANT_QUANTIZATION;
-            if (Btype != GGML_TYPE_F16)
-                return NOT_SUPPORTED;
-            tinyBLAS<0, 16, __m512, __m512, ggml_fp16_t, ggml_fp16_t, TC> tb{
-                k, (const ggml_fp16_t*)A, lda, (const ggml_fp16_t*)B, ldb, C, ldc, ith, nth};
-            tb.matmul(m, n, task);
-            return true;
-#elif (defined(__AVX__) || defined(__AVX2__)) && defined(__F16C__)
-            // if (X86_CHECK(F16C)) {
-            if (k % 8)
-                return NOT_SUPPORTED;
-            if (Btype == GGML_TYPE_F32 && n < 2) {
-                tinyBLAS<0, 8, __m256, __m256, ggml_fp16_t, float, TC> tb{
-                    k, (const ggml_fp16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
-                tb.matmul(m, n, task);
-                return true;
-            }
-            if (Btype == GGML_TYPE_F32)
-                return WANT_QUANTIZATION;
-            if (Btype != GGML_TYPE_F16)
-                return NOT_SUPPORTED;
-            tinyBLAS<0, 8, __m256, __m256, ggml_fp16_t, ggml_fp16_t, TC> tb{
-                k, (const ggml_fp16_t*)A, lda, (const ggml_fp16_t*)B, ldb, C, ldc, ith, nth};
-            tb.matmul(m, n, task);
-            return true;
-            // } else {
-            //     return NOT_SUPPORTED;
-            // }
-#elif defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
-            if (n < 2 && !FLAG_precise)
-                // TODO(jart): Why is ggml_vec_dot_f16_unroll() so fast at matvec?
-                return NOT_SUPPORTED;
-            if (precision == GGML_PREC_F32) {
-                if (k % 4)
-                    return NOT_SUPPORTED;
-                if (Btype != GGML_TYPE_F32)
-                    return NOT_SUPPORTED;
-                tinyBLAS<0, 4, float32x4_t, float32x4_t, ggml_fp16_t, float, TC> tb{
-                    k, (const ggml_fp16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
-                tb.matmul(m, n, task);
-                return true;
-            } else {
-                if (k % 8)
-                    return NOT_SUPPORTED;
-                if (Btype == GGML_TYPE_F32)
-                    return WANT_QUANTIZATION;
-                if (Btype != GGML_TYPE_F16)
-                    return NOT_SUPPORTED;
-                tinyBLAS<0, 8, float16x8_t, float16x8_t, ggml_fp16_t, ggml_fp16_t, TC> tb{
-                    k, (const ggml_fp16_t*)A, lda, (const ggml_fp16_t*)B, ldb, C, ldc, ith, nth};
-                tb.matmul(m, n, task);
-                return true;
-            }
-#elif defined(__ARM_NEON) && !defined(_MSC_VER)
-            if (n < 2 && !FLAG_precise)
-                // TODO(jart): Why is ggml_vec_dot_f16_unroll() so fast at matvec?
-                return NOT_SUPPORTED;
-            if (k % 4)
-                return NOT_SUPPORTED;
-            if (Btype != GGML_TYPE_F32)
-                return NOT_SUPPORTED;
-            tinyBLAS<0, 4, float32x4_t, float32x4_t, ggml_fp16_t, float, TC> tb{
-                k, (const ggml_fp16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
-            tb.matmul(m, n, task);
-            return true;
-#else
-            return NOT_SUPPORTED;
-#endif
-        }
-
-        case GGML_TYPE_Q8_0: {
-            if (Btype == GGML_TYPE_F32)
-                return WANT_QUANTIZATION;
-            if (Btype != GGML_TYPE_Q8_0)
-                return NOT_SUPPORTED;
-#if defined(__AVX2__) || defined(__AVX512F__)
-            tinyBLAS_Q0_AVX2<0, block_q8_0, block_q8_0, TC> tb{
-                k, (const block_q8_0*)A, lda, (const block_q8_0*)B, ldb, C, ldc, ith, nth};
-            tb.matmul(m, n, task);
-            return true;
-#elif defined(__ARM_FEATURE_DOTPROD)
-            tinyBLAS_Q0_ARM<0, block_q8_0, block_q8_0, TC> tb{
-                k, (const block_q8_0*)A, lda, (const block_q8_0*)B, ldb, C, ldc, ith, nth};
-            tb.matmul(m, n, task);
-            return true;
-#else
-            return NOT_SUPPORTED;
-#endif
-        }
-
-        case GGML_TYPE_Q4_0: {
-            if (Btype == GGML_TYPE_F32)
-                return WANT_QUANTIZATION;
-            if (Btype != GGML_TYPE_Q8_0)
-                return NOT_SUPPORTED;
-#if defined(__AVX2__) || defined(__AVX512F__)
-            tinyBLAS_Q0_AVX2<0, block_q4_0, block_q8_0, TC> tb{
-                k, (const block_q4_0*)A, lda, (const block_q8_0*)B, ldb, C, ldc, ith, nth};
-            tb.matmul(m, n, task);
-            return true;
-#elif defined(__ARM_FEATURE_DOTPROD)
-            tinyBLAS_Q0_ARM<0, block_q4_0, block_q8_0, TC> tb{
-                k, (const block_q4_0*)A, lda, (const block_q8_0*)B, ldb, C, ldc, ith, nth};
-            tb.matmul(m, n, task);
-            return true;
-#else
-            return NOT_SUPPORTED;
-#endif
-        }
-
-        default:
-            return NOT_SUPPORTED;
-    }
-
-    (void)m;
-    (void)n;
-    (void)k;
-    (void)A;
-    (void)lda;
-    (void)B;
-    (void)ldb;
-    (void)C;
-    (void)ldc;
-    (void)ith;
-    (void)nth;
-    (void)Atype;
-    (void)Btype;
-    (void)precision;
-}
-
-}  // namespace
-
-/**
- * Performs optimized matrix multiplication on CPU.
- *
- * This subroutine may compute C = Aᵀ * B with column major ordering.
- * Despite its name, this isn't a generalized implementation. Work is
- * only performed when a handwritten kernel is written and available.
- * Otherwise the caller should fall back to a general matmul routine.
- *
- * For example, for single-threaded single-precision GEMM you can say
- *
- *     llamafile_sgemm(m, n, k, A, lda, B, ldb, C, ldc, 0, 1,
- *                     GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32,
- *                     GGML_PREC_DEFAULT);
- *
- * @param m is rows in `A` and `C`
- * @param n is cols in `B` and `C`
- * @param k is cols in `A` and rows in `B`
- * @param A is first input matrix (always transposed)
- * @param lda is row stride of `A`
- * @param B is second input matrix (never transposed)
- * @param ldb is row stride of `B`
- * @param C is input/output array of output matrices
- * @param ldc is row stride of `C`
- * @param ith is thread id (must be less than `nth`)
- * @param nth is number of threads (must be greater than zero)
- * @param Atype is GGML data type of `A`
- * @param Btype is GGML data type of `B`
- * @param Ctype is GGML data type of `C`
- * @param precision may be used to control the internal compute type
- * @return true if this function was able to service the matmul request
- */
-bool llamafile_sgemm(long m, long n, long k, const void* A, long lda, const void* B, long ldb, void* C, long ldc, int ith, int nth, int task, int Atype, int Btype, int Ctype, int precision) {
-    assert(m >= 0);
-    assert(n >= 0);
-    assert(k >= 0);
-    assert(lda >= k);
-    assert(ldb >= k);
-    assert(ldc >= m);
-    assert(nth > 0);
-    assert(ith < nth);
-
-#if QK_K == 256
-#if defined(__x86_64__) || defined(_M_X64)
-#if defined(__AVX2__) && (defined(__FMA__) || (defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))))
-    /* 
-    moonll
-    more Btype accept
-    }*/
-
-    if (Ctype == GGML_TYPE_F32){
-        if (iqk_mul_mat(m, n, k * ggml_blck_size(ggml_type(Atype)), Atype, A,lda,Btype, B,ldb, (float*)C, ldc, ith, nth)) {
-            return true;
-        }
-    }
-
-#endif
-#elif defined __aarch64__ && defined __ARM_FEATURE_DOTPROD && !defined _MSC_VER
-    if (Btype == GGML_TYPE_Q8_K && Ctype == GGML_TYPE_F32) {
-        if (iqk_mul_mat(m, n, k * QK_K, Atype, A, B, (float*)C, ldc, ith, nth)) {
-            return true;
-        }
-    }
-    if ((Btype == GGML_TYPE_Q8_0 || Btype == GGML_TYPE_Q8_1) && Ctype == GGML_TYPE_F32) {
-        // assert(QK8_0 == QK8_1 == QK4_0 == QK4_1 == QK5_0 == QK5_1 == 32);
-        assert((QK8_0 == 32) && (QK8_1 == 32) && (QK4_0 == 32) && (QK4_1 == 32) && (QK5_0 == 32) && (QK5_1 == 32));
-        if (iqk_mul_mat(m, n, k * QK8_0, Atype, A, B, (float*)C, ldc, ith, nth)) {
-            return true;
-        }
-    }
-#endif
-#endif
-
-    switch (Ctype) {
-        case GGML_TYPE_F32:
-            return llamafile_sgemm_impl(m, n, k, A, lda, B, ldb, (float*)C, ldc, ith, nth, task, Atype,
-                                        Btype, Ctype, precision);
-        default:
-            return NOT_SUPPORTED;
-    }
-}
+    // 使用 ARM 版本
+    #include "tinyblas_cpu_sgemm_x86.inc"
+#endif

third_party/llamafile/tinyblas_cpu_sgemm_arm.inc

@@ -0,0 +1,471 @@
+// Adapted from
+// https://github.com/Mozilla-Ocho/llamafile/blob/0.8.8/llamafile/tinyblas_cpu_sgemm.inc
+// Copyrigth 2024 Mozilla Foundation.
+// Copyright(c) 2024 by KVCache.AI, All Rights Reserved.
+
+// -*- mode:c++;indent-tabs-mode:nil;c-basic-offset:4;coding:utf-8 -*-
+// vi: set et ft=cpp ts=4 sts=4 sw=4 fenc=utf-8 :vi
+//
+// Copyright 2024 Mozilla Foundation
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "tinyblas_cpu.h"
+#include <arm_neon.h>
+#include <ostream>
+#include <iostream>
+//
+//
+//                                ██████╗ ██╗   █████╗ ██████╗
+//         ██████╗██╗██╗ ██╗██═██╗██╔══██╗██║  ██╔══██╗██╔═══╝
+//         ╚═██╔═╝██║███▄██║██ ██║██████╔╝██║  ███████║██████╗
+//           ██║  ██║██▀███║╚███╔╝██╔══██╗██║  ██╔══██║╔═══██║
+//           ██║  ██║██║ ██║ ███║ ██████╔╝████╗██║  ██║██████║
+//           ╚═╝  ╚═╝╚═╝ ╚═╝ ╚══╝ ╚═════╝ ╚═══╝╚═╝  ╚═╝╚═════╝
+//
+//                   BASIC LINEAR ALGEBRA SUBPROGRAMS
+//
+//
+// This file implements multithreaded CPU matrix multiplication for the
+// common contiguous use case C = Aᵀ * B. These kernels are designed to
+// have excellent performance[1] for matrices that fit in the CPU cache
+// without imposing any overhead such as cache filling or malloc calls.
+//
+// This implementation does not guarantee any upper bound with rounding
+// errors, which grow along with k. Our goal's to maximally exploit the
+// hardware for performance, and then use whatever resources remain for
+// improving numerical accuracy.
+//
+// [1] J. Tunney, ‘LLaMA Now Goes Faster on CPUs’, Mar. 2024. [Online].
+//     Available: https://justine.lol/matmul/. [Accessed: 29-Mar-2024].
+
+namespace {
+
+template <typename TC>
+void SgemmHelperN1Neon2(long m, long n, long k, const float16_t* A, long lda, const float16_t* B, long ldb,
+                        TC* C, long ldc, int ith, int nth) {
+    // A m * k    B n * k    c n * m
+    const long NVL = 8;
+    long kk = k / (NVL * 4);
+    kk = kk * (NVL * 4);
+    long length = (m / nth) + (ith < (m % nth) ? 1 : 0);
+    long startRow = ith * (m / nth) + (ith < (m % nth) ? ith : (m % nth));
+    long endRow = startRow + length;
+    for (long i = startRow; i < endRow; i ++) {
+        const float16_t* tA = A + i * lda;
+        float32x4_t c0 = vdupq_n_f32(0);
+        float32x4_t c1 = vdupq_n_f32(0);
+        float32x4_t c2 = vdupq_n_f32(0);
+        float32x4_t c3 = vdupq_n_f32(0);
+        float32x4_t c4 = vdupq_n_f32(0);
+        float32x4_t c5 = vdupq_n_f32(0);
+        float32x4_t c6 = vdupq_n_f32(0);
+        float32x4_t c7 = vdupq_n_f32(0);
+        for (long j = 0; j < kk; j += NVL * 4) {
+            __builtin_prefetch(tA + 192, 0, 0);
+            float16x8_t a0 = vld1q_f16(tA + j);
+            float16x8_t b0 = vld1q_f16(B + j);
+            c0 = vfmlalq_low_f16(c0, a0, b0);
+            c1 = vfmlalq_high_f16(c1, a0, b0);
+            float16x8_t a1 = vld1q_f16(tA + j + NVL);
+            float16x8_t b1 = vld1q_f16(B + j + NVL);
+            c2 = vfmlalq_low_f16(c2, a1, b1);
+            c3 = vfmlalq_high_f16(c3, a1, b1);
+            float16x8_t a2 = vld1q_f16(tA + j + NVL * 2);
+            float16x8_t b2 = vld1q_f16(B + j + NVL * 2);
+            c4 = vfmlalq_low_f16(c4, a2, b2);
+            c5 = vfmlalq_high_f16(c5, a2, b2);
+            float16x8_t a3 = vld1q_f16(tA + j + NVL * 3);
+            float16x8_t b3 = vld1q_f16(B + j + NVL * 3);
+            c6 = vfmlalq_low_f16(c6, a3, b3);
+            c7 = vfmlalq_high_f16(c7, a3, b3);
+        }
+        if (k - kk >= NVL * 2) {
+            float16x8_t a0 = vld1q_f16(tA + kk);
+            float16x8_t b0 = vld1q_f16(B + kk);
+            c0 = vfmlalq_low_f16(c0, a0, b0);
+            c1 = vfmlalq_high_f16(c1, a0, b0);
+            float16x8_t a1 = vld1q_f16(tA + kk + NVL);
+            float16x8_t b1 = vld1q_f16(B + kk + NVL);
+            c2 = vfmlalq_low_f16(c2, a1, b1);
+            c3 = vfmlalq_high_f16(c3, a1, b1);
+            kk += NVL * 2;
+        }
+        if (k - kk >= NVL) {
+            float16x8_t a = vld1q_f16(tA + kk);
+            float16x8_t b = vld1q_f16(B + kk);
+            c0 = vfmlalq_low_f16(c0, a, b);
+            c1 = vfmlalq_high_f16(c1, a, b);
+            kk += NVL;
+        }
+        TC sum = 0.0f;
+        for (long j = kk; j < k; j ++) {
+            sum += (float32_t)tA[j] * (float32_t)B[j];
+        }
+        c0 = vaddq_f32(c0, c1);
+        c2 = vaddq_f32(c2, c3);
+        c4 = vaddq_f32(c4, c5);
+        c6 = vaddq_f32(c6, c7);
+        c0 = vaddq_f32(c0, c2);
+        c4 = vaddq_f32(c4, c6);
+        sum += vaddvq_f32(c0) + vaddvq_f32(c4);
+        C[i] = sum;
+    }
+    return;
+}
+
+template <typename TC>
+void SgemmHelperN1(long m, long n, long k, const ggml_fp16_t* A_, long lda, const ggml_fp16_t* B_, long ldb,
+                   TC* C, long ldc, int ith, int nth) {
+    // A m * k    B n * k    c n * m
+    float16_t *A = (float16_t*)A_;
+    float16_t *B = (float16_t*)B_;
+    long rowsPerThread = m / nth;
+    long startRow = ith * rowsPerThread;
+    long endRow = (ith == nth - 1) ? m : startRow + rowsPerThread;
+    for (long i = startRow; i < endRow; i ++) {
+        TC sum = 0.0f;
+        for (long j = 0; j < k; j ++) {
+            sum += (float32_t)A[i * lda + j] * (float32_t)B[j];
+        }
+        C[i] = sum;
+    }
+    return;
+}
+
+template <typename TC>
+bool llamafile_sgemm_impl(long m, long n, long k, const void* A, long lda, const void* B, long ldb, TC* C, long ldc, int ith, int nth, int task, int Atype, int Btype, int Ctype, int precision) {
+    switch (Atype) {
+        case GGML_TYPE_F32: {
+            if (Btype != GGML_TYPE_F32)
+                return NOT_SUPPORTED;
+#if defined(__AVX512F__)
+            if (k % 16)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 16, __m512, __m512, float, float, TC> tb{
+                k, (const float*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#elif defined(__AVX__) || defined(__AVX2__)
+            if (k % 8)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 8, __m256, __m256, float, float, TC> tb{
+                k, (const float*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#elif defined(__ARM_NEON)
+            if (k % 4)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 4, float32x4_t, float32x4_t, float, float, TC> tb{
+                k, (const float*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#else
+            return NOT_SUPPORTED;
+#endif
+        }
+
+        case GGML_TYPE_BF16: {
+#if defined(__AVX512BF16__)
+            if (k % 32)
+                return NOT_SUPPORTED;
+            if (Btype == GGML_TYPE_F32 && n < 2) {
+                tinyBLAS<0, 16, __m512, __m512, ggml_bf16_t, float, TC> tb{
+                    k, (const ggml_bf16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+                tb.matmul(m, n, task);
+                return true;
+            }
+            if (Btype == GGML_TYPE_F32)
+                return WANT_QUANTIZATION;
+            if (Btype != GGML_TYPE_BF16)
+                return NOT_SUPPORTED;
+            if (!FLAG_precise) {
+                tinyBLAS<0, 32, __m512, __m512bh, ggml_bf16_t, ggml_bf16_t, TC> tb{
+                    k, (const ggml_bf16_t*)A, lda, (const ggml_bf16_t*)B, ldb, C, ldc, ith, nth};
+                tb.matmul(m, n, task);
+                return true;
+            } else {
+                tinyBLAS<0, 16, __m512, __m512, ggml_bf16_t, ggml_bf16_t, TC> tb{
+                    k, (const ggml_bf16_t*)A, lda, (const ggml_bf16_t*)B, ldb, C, ldc, ith, nth};
+                tb.matmul(m, n, task);
+                return true;
+            }
+#elif defined(__AVX512F__)
+            if (k % 16)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 16, __m512, __m512, ggml_bf16_t, float, TC> tb{
+                k, (const ggml_bf16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#elif defined(__AVX2__)
+            if (k % 8)
+                return NOT_SUPPORTED;
+            if (Btype != GGML_TYPE_F32)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 8, __m256, __m256, ggml_bf16_t, float, TC> tb{
+                k, (const ggml_bf16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#elif defined(__ARM_NEON) && !defined(_MSC_VER)
+            if (k % 4)
+                return NOT_SUPPORTED;
+            if (Btype != GGML_TYPE_F32)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 4, float32x4_t, float32x4_t, ggml_bf16_t, float, TC> tb{
+                k, (const ggml_bf16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#else
+            return NOT_SUPPORTED;
+#endif
+        }
+
+        case GGML_TYPE_F16: {
+#if defined(__AVX512F__)
+            if (k % 16)
+                return NOT_SUPPORTED;
+            if (Btype == GGML_TYPE_F32 && n < 2) {
+                tinyBLAS<0, 16, __m512, __m512, ggml_fp16_t, float, TC> tb{
+                    k, (const ggml_fp16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+                tb.matmul(m, n, task);
+                return true;
+            }
+            if (Btype == GGML_TYPE_F32)
+                return WANT_QUANTIZATION;
+            if (Btype != GGML_TYPE_F16)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 16, __m512, __m512, ggml_fp16_t, ggml_fp16_t, TC> tb{
+                k, (const ggml_fp16_t*)A, lda, (const ggml_fp16_t*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#elif (defined(__AVX__) || defined(__AVX2__)) && defined(__F16C__)
+            // if (X86_CHECK(F16C)) {
+            if (k % 8)
+                return NOT_SUPPORTED;
+            if (Btype == GGML_TYPE_F32 && n < 2) {
+                tinyBLAS<0, 8, __m256, __m256, ggml_fp16_t, float, TC> tb{
+                    k, (const ggml_fp16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+                tb.matmul(m, n, task);
+                return true;
+            }
+            if (Btype == GGML_TYPE_F32)
+                return WANT_QUANTIZATION;
+            if (Btype != GGML_TYPE_F16)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 8, __m256, __m256, ggml_fp16_t, ggml_fp16_t, TC> tb{
+                k, (const ggml_fp16_t*)A, lda, (const ggml_fp16_t*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+            // } else {
+            //     return NOT_SUPPORTED;
+            // }
+#elif defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
+            if (n < 2 && !FLAG_precise) {
+                // TODO(jart): Why is ggml_vec_dot_f16_unroll() so fast at matvec?
+                if (Btype == GGML_TYPE_F16 && task == GGML_TASK_TYPE_COMPUTE) {
+                    SgemmHelperN1Neon2<TC>(m, n, k, (const float16_t*)A, lda, (const float16_t*)B, ldb, C, ldc, ith, nth);
+                    return true;
+                }
+                return NOT_SUPPORTED;
+            }
+            if (precision == GGML_PREC_F32) {
+                if (k % 4)
+                    return NOT_SUPPORTED;
+                if (Btype != GGML_TYPE_F32)
+                    return NOT_SUPPORTED;
+                tinyBLAS<0, 4, float32x4_t, float32x4_t, ggml_fp16_t, float, TC> tb{
+                    k, (const ggml_fp16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+                tb.matmul(m, n, task);
+                return true;
+            } else {
+                if (k % 8)
+                    return NOT_SUPPORTED;
+                if (Btype == GGML_TYPE_F32)
+                    return WANT_QUANTIZATION;
+                if (Btype != GGML_TYPE_F16)
+                    return NOT_SUPPORTED;
+                tinyBLAS<0, 8, float16x8_t, float16x8_t, ggml_fp16_t, ggml_fp16_t, TC> tb{
+                    k, (const ggml_fp16_t*)A, lda, (const ggml_fp16_t*)B, ldb, C, ldc, ith, nth};
+                tb.matmul(m, n, task);
+                return true;
+            }
+#elif defined(__ARM_NEON) && !defined(_MSC_VER)
+            if (n < 2 && !FLAG_precise) {
+                // TODO(jart): Why is ggml_vec_dot_f16_unroll() so fast at matvec?
+                if (Btype == GGML_TYPE_F16 && task == GGML_TASK_TYPE_COMPUTE) {
+                    SgemmHelperN1Neon2<TC>(m, n, k, (const float16_t*)A, lda, (const float16_t*)B, ldb, C, ldc, ith, nth);
+                    return true;
+                }
+                return NOT_SUPPORTED;
+            }
+            if (k % 4)
+                return NOT_SUPPORTED;
+            if (Btype != GGML_TYPE_F32)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 4, float32x4_t, float32x4_t, ggml_fp16_t, float, TC> tb{
+                k, (const ggml_fp16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#else
+            return NOT_SUPPORTED;
+#endif
+        }
+
+        case GGML_TYPE_Q8_0: {
+            if (Btype == GGML_TYPE_F32)
+                return WANT_QUANTIZATION;
+            if (Btype != GGML_TYPE_Q8_0)
+                return NOT_SUPPORTED;
+#if defined(__AVX2__) || defined(__AVX512F__)
+            tinyBLAS_Q0_AVX2<0, block_q8_0, block_q8_0, TC> tb{
+                k, (const block_q8_0*)A, lda, (const block_q8_0*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#elif defined(__ARM_FEATURE_DOTPROD)
+            tinyBLAS_Q0_ARM<0, block_q8_0, block_q8_0, TC> tb{
+                k, (const block_q8_0*)A, lda, (const block_q8_0*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#else
+            return NOT_SUPPORTED;
+#endif
+        }
+
+        case GGML_TYPE_Q4_0: {
+            if (Btype == GGML_TYPE_F32)
+                return WANT_QUANTIZATION;
+            if (Btype != GGML_TYPE_Q8_0)
+                return NOT_SUPPORTED;
+#if defined(__AVX2__) || defined(__AVX512F__)
+            tinyBLAS_Q0_AVX2<0, block_q4_0, block_q8_0, TC> tb{
+                k, (const block_q4_0*)A, lda, (const block_q8_0*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#elif defined(__ARM_FEATURE_DOTPROD)
+            tinyBLAS_Q0_ARM<0, block_q4_0, block_q8_0, TC> tb{
+                k, (const block_q4_0*)A, lda, (const block_q8_0*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#else
+            return NOT_SUPPORTED;
+#endif
+        }
+
+        default:
+            return NOT_SUPPORTED;
+    }
+
+    (void)m;
+    (void)n;
+    (void)k;
+    (void)A;
+    (void)lda;
+    (void)B;
+    (void)ldb;
+    (void)C;
+    (void)ldc;
+    (void)ith;
+    (void)nth;
+    (void)Atype;
+    (void)Btype;
+    (void)precision;
+}
+
+}  // namespace
+
+/**
+ * Performs optimized matrix multiplication on CPU.
+ *
+ * This subroutine may compute C = Aᵀ * B with column major ordering.
+ * Despite its name, this isn't a generalized implementation. Work is
+ * only performed when a handwritten kernel is written and available.
+ * Otherwise the caller should fall back to a general matmul routine.
+ *
+ * For example, for single-threaded single-precision GEMM you can say
+ *
+ *     llamafile_sgemm(m, n, k, A, lda, B, ldb, C, ldc, 0, 1,
+ *                     GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32,
+ *                     GGML_PREC_DEFAULT);
+ *
+ * @param m is rows in `A` and `C`
+ * @param n is cols in `B` and `C`
+ * @param k is cols in `A` and rows in `B`
+ * @param A is first input matrix (always transposed)
+ * @param lda is row stride of `A`
+ * @param B is second input matrix (never transposed)
+ * @param ldb is row stride of `B`
+ * @param C is input/output array of output matrices
+ * @param ldc is row stride of `C`
+ * @param ith is thread id (must be less than `nth`)
+ * @param nth is number of threads (must be greater than zero)
+ * @param Atype is GGML data type of `A`
+ * @param Btype is GGML data type of `B`
+ * @param Ctype is GGML data type of `C`
+ * @param precision may be used to control the internal compute type
+ * @return true if this function was able to service the matmul request
+ */
+bool llamafile_sgemm(long m, long n, long k, const void* A, long lda, const void* B, long ldb, void* C, long ldc, int ith, int nth, int task, int Atype, int Btype, int Ctype, int precision) {
+    assert(m >= 0);
+    assert(n >= 0);
+    assert(k >= 0);
+    assert(lda >= k);
+    assert(ldb >= k);
+    assert(ldc >= m);
+    assert(nth > 0);
+    assert(ith < nth);
+
+#if QK_K == 256
+#if defined(__x86_64__) || defined(_M_X64)
+#if defined(__AVX2__) && (defined(__FMA__) || (defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))))
+    /* 
+    moonll
+    more Btype accept
+    }*/
+    // if (X86_CHECK(AVX2) && X86_CHECK(FMA)) {
+    if (Btype == GGML_TYPE_Q8_K && Ctype == GGML_TYPE_F32){
+        if (iqk_mul_mat(m, n, k * QK_K, Atype, A, B, (float*)C, ldc, ith, nth)) {
+            return true;
+        }
+    }
+    if ((Btype == GGML_TYPE_Q8_0 || Btype == GGML_TYPE_Q8_1) && Ctype == GGML_TYPE_F32) {
+        assert((QK8_0 == 32) && (QK8_1 == 32) && (QK4_0 == 32) && (QK4_1 == 32) && (QK5_0 == 32) && (QK5_1 == 32));
+        if (iqk_mul_mat(m, n, k * QK8_0, Atype, A, B, (float*)C, ldc, ith, nth)) {
+            return true;
+        }
+    }
+
+#endif
+#elif defined __aarch64__ && defined __ARM_FEATURE_DOTPROD && !defined _MSC_VER
+    if (Btype == GGML_TYPE_Q8_K && Ctype == GGML_TYPE_F32) {
+        if (iqk_mul_mat(m, n, k * QK_K, Atype, A, k, Btype, B, k, (float*)C, ldc, ith, nth)) {
+            return true;
+        }
+    }
+    if ((Btype == GGML_TYPE_Q8_0 || Btype == GGML_TYPE_Q8_1) && Ctype == GGML_TYPE_F32) {
+        // assert(QK8_0 == QK8_1 == QK4_0 == QK4_1 == QK5_0 == QK5_1 == 32);
+        assert((QK8_0 == 32) && (QK8_1 == 32) && (QK4_0 == 32) && (QK4_1 == 32) && (QK5_0 == 32) && (QK5_1 == 32));
+        if (iqk_mul_mat(m, n, k * QK8_0, Atype, A, k, Btype, B, k, (float*)C, ldc, ith, nth)) {
+            return true;
+        }
+    }
+#endif
+#endif
+
+    switch (Ctype) {
+        case GGML_TYPE_F32:
+            return llamafile_sgemm_impl(m, n, k, A, lda, B, ldb, (float*)C, ldc, ith, nth, task, Atype,
+                                        Btype, Ctype, precision);
+        default:
+            return NOT_SUPPORTED;
+    }
+}

third_party/llamafile/tinyblas_cpu_sgemm_x86.inc

@@ -0,0 +1,361 @@
+// Adapted from
+// https://github.com/Mozilla-Ocho/llamafile/blob/0.8.8/llamafile/tinyblas_cpu_sgemm.inc
+// Copyrigth 2024 Mozilla Foundation.
+// Copyright(c) 2024 by KVCache.AI, All Rights Reserved.
+
+// -*- mode:c++;indent-tabs-mode:nil;c-basic-offset:4;coding:utf-8 -*-
+// vi: set et ft=cpp ts=4 sts=4 sw=4 fenc=utf-8 :vi
+//
+// Copyright 2024 Mozilla Foundation
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "tinyblas_cpu.h"
+
+//
+//
+//                                ██████╗ ██╗   █████╗ ██████╗
+//         ██████╗██╗██╗ ██╗██═██╗██╔══██╗██║  ██╔══██╗██╔═══╝
+//         ╚═██╔═╝██║███▄██║██ ██║██████╔╝██║  ███████║██████╗
+//           ██║  ██║██▀███║╚███╔╝██╔══██╗██║  ██╔══██║╔═══██║
+//           ██║  ██║██║ ██║ ███║ ██████╔╝████╗██║  ██║██████║
+//           ╚═╝  ╚═╝╚═╝ ╚═╝ ╚══╝ ╚═════╝ ╚═══╝╚═╝  ╚═╝╚═════╝
+//
+//                   BASIC LINEAR ALGEBRA SUBPROGRAMS
+//
+//
+// This file implements multithreaded CPU matrix multiplication for the
+// common contiguous use case C = Aᵀ * B. These kernels are designed to
+// have excellent performance[1] for matrices that fit in the CPU cache
+// without imposing any overhead such as cache filling or malloc calls.
+//
+// This implementation does not guarantee any upper bound with rounding
+// errors, which grow along with k. Our goal's to maximally exploit the
+// hardware for performance, and then use whatever resources remain for
+// improving numerical accuracy.
+//
+// [1] J. Tunney, ‘LLaMA Now Goes Faster on CPUs’, Mar. 2024. [Online].
+//     Available: https://justine.lol/matmul/. [Accessed: 29-Mar-2024].
+
+namespace {
+
+template <typename TC>
+bool llamafile_sgemm_impl(long m, long n, long k, const void* A, long lda, const void* B, long ldb, TC* C, long ldc, int ith, int nth, int task, int Atype, int Btype, int Ctype, int precision) {
+    switch (Atype) {
+        case GGML_TYPE_F32: {
+            if (Btype != GGML_TYPE_F32)
+                return NOT_SUPPORTED;
+#if defined(__AVX512F__)
+            if (k % 16)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 16, __m512, __m512, float, float, TC> tb{
+                k, (const float*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#elif defined(__AVX__) || defined(__AVX2__)
+            if (k % 8)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 8, __m256, __m256, float, float, TC> tb{
+                k, (const float*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#elif defined(__ARM_NEON)
+            if (k % 4)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 4, float32x4_t, float32x4_t, float, float, TC> tb{
+                k, (const float*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#else
+            return NOT_SUPPORTED;
+#endif
+        }
+
+        case GGML_TYPE_BF16: {
+#if defined(__AVX512BF16__)
+            if (k % 32)
+                return NOT_SUPPORTED;
+            if (Btype == GGML_TYPE_F32 && n < 2) {
+                tinyBLAS<0, 16, __m512, __m512, ggml_bf16_t, float, TC> tb{
+                    k, (const ggml_bf16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+                tb.matmul(m, n, task);
+                return true;
+            }
+            if (Btype == GGML_TYPE_F32)
+                return WANT_QUANTIZATION;
+            if (Btype != GGML_TYPE_BF16)
+                return NOT_SUPPORTED;
+            if (!FLAG_precise) {
+                tinyBLAS<0, 32, __m512, __m512bh, ggml_bf16_t, ggml_bf16_t, TC> tb{
+                    k, (const ggml_bf16_t*)A, lda, (const ggml_bf16_t*)B, ldb, C, ldc, ith, nth};
+                tb.matmul(m, n, task);
+                return true;
+            } else {
+                tinyBLAS<0, 16, __m512, __m512, ggml_bf16_t, ggml_bf16_t, TC> tb{
+                    k, (const ggml_bf16_t*)A, lda, (const ggml_bf16_t*)B, ldb, C, ldc, ith, nth};
+                tb.matmul(m, n, task);
+                return true;
+            }
+#elif defined(__AVX512F__)
+            if (k % 16)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 16, __m512, __m512, ggml_bf16_t, float, TC> tb{
+                k, (const ggml_bf16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#elif defined(__AVX2__)
+            if (k % 8)
+                return NOT_SUPPORTED;
+            if (Btype != GGML_TYPE_F32)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 8, __m256, __m256, ggml_bf16_t, float, TC> tb{
+                k, (const ggml_bf16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#elif defined(__ARM_NEON) && !defined(_MSC_VER)
+            if (k % 4)
+                return NOT_SUPPORTED;
+            if (Btype != GGML_TYPE_F32)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 4, float32x4_t, float32x4_t, ggml_bf16_t, float, TC> tb{
+                k, (const ggml_bf16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#else
+            return NOT_SUPPORTED;
+#endif
+        }
+
+        case GGML_TYPE_F16: {
+#if defined(__AVX512F__)
+            if (k % 16)
+                return NOT_SUPPORTED;
+            if (Btype == GGML_TYPE_F32 && n < 2) {
+                tinyBLAS<0, 16, __m512, __m512, ggml_fp16_t, float, TC> tb{
+                    k, (const ggml_fp16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+                tb.matmul(m, n, task);
+                return true;
+            }
+            if (Btype == GGML_TYPE_F32)
+                return WANT_QUANTIZATION;
+            if (Btype != GGML_TYPE_F16)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 16, __m512, __m512, ggml_fp16_t, ggml_fp16_t, TC> tb{
+                k, (const ggml_fp16_t*)A, lda, (const ggml_fp16_t*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#elif (defined(__AVX__) || defined(__AVX2__)) && defined(__F16C__)
+            // if (X86_CHECK(F16C)) {
+            if (k % 8)
+                return NOT_SUPPORTED;
+            if (Btype == GGML_TYPE_F32 && n < 2) {
+                tinyBLAS<0, 8, __m256, __m256, ggml_fp16_t, float, TC> tb{
+                    k, (const ggml_fp16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+                tb.matmul(m, n, task);
+                return true;
+            }
+            if (Btype == GGML_TYPE_F32)
+                return WANT_QUANTIZATION;
+            if (Btype != GGML_TYPE_F16)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 8, __m256, __m256, ggml_fp16_t, ggml_fp16_t, TC> tb{
+                k, (const ggml_fp16_t*)A, lda, (const ggml_fp16_t*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+            // } else {
+            //     return NOT_SUPPORTED;
+            // }
+#elif defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
+            if (n < 2 && !FLAG_precise)
+                // TODO(jart): Why is ggml_vec_dot_f16_unroll() so fast at matvec?
+                return NOT_SUPPORTED;
+            if (precision == GGML_PREC_F32) {
+                if (k % 4)
+                    return NOT_SUPPORTED;
+                if (Btype != GGML_TYPE_F32)
+                    return NOT_SUPPORTED;
+                tinyBLAS<0, 4, float32x4_t, float32x4_t, ggml_fp16_t, float, TC> tb{
+                    k, (const ggml_fp16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+                tb.matmul(m, n, task);
+                return true;
+            } else {
+                if (k % 8)
+                    return NOT_SUPPORTED;
+                if (Btype == GGML_TYPE_F32)
+                    return WANT_QUANTIZATION;
+                if (Btype != GGML_TYPE_F16)
+                    return NOT_SUPPORTED;
+                tinyBLAS<0, 8, float16x8_t, float16x8_t, ggml_fp16_t, ggml_fp16_t, TC> tb{
+                    k, (const ggml_fp16_t*)A, lda, (const ggml_fp16_t*)B, ldb, C, ldc, ith, nth};
+                tb.matmul(m, n, task);
+                return true;
+            }
+#elif defined(__ARM_NEON) && !defined(_MSC_VER)
+            if (n < 2 && !FLAG_precise)
+                // TODO(jart): Why is ggml_vec_dot_f16_unroll() so fast at matvec?
+                return NOT_SUPPORTED;
+            if (k % 4)
+                return NOT_SUPPORTED;
+            if (Btype != GGML_TYPE_F32)
+                return NOT_SUPPORTED;
+            tinyBLAS<0, 4, float32x4_t, float32x4_t, ggml_fp16_t, float, TC> tb{
+                k, (const ggml_fp16_t*)A, lda, (const float*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#else
+            return NOT_SUPPORTED;
+#endif
+        }
+
+        case GGML_TYPE_Q8_0: {
+            if (Btype == GGML_TYPE_F32)
+                return WANT_QUANTIZATION;
+            if (Btype != GGML_TYPE_Q8_0)
+                return NOT_SUPPORTED;
+#if defined(__AVX2__) || defined(__AVX512F__)
+            tinyBLAS_Q0_AVX2<0, block_q8_0, block_q8_0, TC> tb{
+                k, (const block_q8_0*)A, lda, (const block_q8_0*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#elif defined(__ARM_FEATURE_DOTPROD)
+            tinyBLAS_Q0_ARM<0, block_q8_0, block_q8_0, TC> tb{
+                k, (const block_q8_0*)A, lda, (const block_q8_0*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#else
+            return NOT_SUPPORTED;
+#endif
+        }
+
+        case GGML_TYPE_Q4_0: {
+            if (Btype == GGML_TYPE_F32)
+                return WANT_QUANTIZATION;
+            if (Btype != GGML_TYPE_Q8_0)
+                return NOT_SUPPORTED;
+#if defined(__AVX2__) || defined(__AVX512F__)
+            tinyBLAS_Q0_AVX2<0, block_q4_0, block_q8_0, TC> tb{
+                k, (const block_q4_0*)A, lda, (const block_q8_0*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#elif defined(__ARM_FEATURE_DOTPROD)
+            tinyBLAS_Q0_ARM<0, block_q4_0, block_q8_0, TC> tb{
+                k, (const block_q4_0*)A, lda, (const block_q8_0*)B, ldb, C, ldc, ith, nth};
+            tb.matmul(m, n, task);
+            return true;
+#else
+            return NOT_SUPPORTED;
+#endif
+        }
+
+        default:
+            return NOT_SUPPORTED;
+    }
+
+    (void)m;
+    (void)n;
+    (void)k;
+    (void)A;
+    (void)lda;
+    (void)B;
+    (void)ldb;
+    (void)C;
+    (void)ldc;
+    (void)ith;
+    (void)nth;
+    (void)Atype;
+    (void)Btype;
+    (void)precision;
+}
+
+}  // namespace
+
+/**
+ * Performs optimized matrix multiplication on CPU.
+ *
+ * This subroutine may compute C = Aᵀ * B with column major ordering.
+ * Despite its name, this isn't a generalized implementation. Work is
+ * only performed when a handwritten kernel is written and available.
+ * Otherwise the caller should fall back to a general matmul routine.
+ *
+ * For example, for single-threaded single-precision GEMM you can say
+ *
+ *     llamafile_sgemm(m, n, k, A, lda, B, ldb, C, ldc, 0, 1,
+ *                     GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32,
+ *                     GGML_PREC_DEFAULT);
+ *
+ * @param m is rows in `A` and `C`
+ * @param n is cols in `B` and `C`
+ * @param k is cols in `A` and rows in `B`
+ * @param A is first input matrix (always transposed)
+ * @param lda is row stride of `A`
+ * @param B is second input matrix (never transposed)
+ * @param ldb is row stride of `B`
+ * @param C is input/output array of output matrices
+ * @param ldc is row stride of `C`
+ * @param ith is thread id (must be less than `nth`)
+ * @param nth is number of threads (must be greater than zero)
+ * @param Atype is GGML data type of `A`
+ * @param Btype is GGML data type of `B`
+ * @param Ctype is GGML data type of `C`
+ * @param precision may be used to control the internal compute type
+ * @return true if this function was able to service the matmul request
+ */
+bool llamafile_sgemm(long m, long n, long k, const void* A, long lda, const void* B, long ldb, void* C, long ldc, int ith, int nth, int task, int Atype, int Btype, int Ctype, int precision) {
+    assert(m >= 0);
+    assert(n >= 0);
+    assert(k >= 0);
+    assert(lda >= k);
+    assert(ldb >= k);
+    assert(ldc >= m);
+    assert(nth > 0);
+    assert(ith < nth);
+
+#if QK_K == 256
+#if defined(__x86_64__) || defined(_M_X64)
+#if defined(__AVX2__) && (defined(__FMA__) || (defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))))
+    /* 
+    moonll
+    more Btype accept
+    }*/
+
+    if (Ctype == GGML_TYPE_F32){
+        if (iqk_mul_mat(m, n, k * ggml_blck_size(ggml_type(Atype)), Atype, A,lda,Btype, B,ldb, (float*)C, ldc, ith, nth)) {
+            return true;
+        }
+    }
+
+#endif
+#elif defined __aarch64__ && defined __ARM_FEATURE_DOTPROD && !defined _MSC_VER
+    if (Btype == GGML_TYPE_Q8_K && Ctype == GGML_TYPE_F32) {
+        if (iqk_mul_mat(m, n, k * QK_K, Atype, A, B, (float*)C, ldc, ith, nth)) {
+            return true;
+        }
+    }
+    if ((Btype == GGML_TYPE_Q8_0 || Btype == GGML_TYPE_Q8_1) && Ctype == GGML_TYPE_F32) {
+        // assert(QK8_0 == QK8_1 == QK4_0 == QK4_1 == QK5_0 == QK5_1 == 32);
+        assert((QK8_0 == 32) && (QK8_1 == 32) && (QK4_0 == 32) && (QK4_1 == 32) && (QK5_0 == 32) && (QK5_1 == 32));
+        if (iqk_mul_mat(m, n, k * QK8_0, Atype, A, B, (float*)C, ldc, ith, nth)) {
+            return true;
+        }
+    }
+#endif
+#endif
+
+    switch (Ctype) {
+        case GGML_TYPE_F32:
+            return llamafile_sgemm_impl(m, n, k, A, lda, B, ldb, (float*)C, ldc, ith, nth, task, Atype,
+                                        Btype, Ctype, precision);
+        default:
+            return NOT_SUPPORTED;
+    }
+}