vrr/kvcache-ai-ktransformers
2
0
Fork 0
mirror of https://github.com/kvcache-ai/ktransformers.git synced 2025-09-10 15:29:39 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

merge main; Add torch q8 linear

Browse source
This commit is contained in:
Azure-Tang 2025-03-14 05:52:07 -04:00
parent 6c4ed59175
commit ed8437413b
27 changed files with 1561 additions and 114 deletions
Download patch file Download diff file Expand all files Collapse all files

3
ktransformers/local_chat.py
View file

@ -31,6 +31,7 @@ from ktransformers.models.modeling_mixtral import MixtralForCausalLM
from ktransformers.util.utils import prefill_and_generate, get_compute_capability
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,
@ -169,7 +170,7 @@ def local_chat(
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:
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_prefill_size = chunk_prefill_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