vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 01:24:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	tests/test-backend-ops.cpp
This commit is contained in:
Concedo 2024-05-22 15:04:31 +08:00
commit 4a3c8c190b
25 changed files with 829 additions and 1596 deletions
Download patch file Download diff file Expand all files Collapse all files

32
common/common.cpp
View file

@ -1355,7 +1355,12 @@ void gpt_params_handle_model_default(gpt_params & params) {
} }
params.hf_file = params.model; params.hf_file = params.model;
} else if (params.model.empty()) { } else if (params.model.empty()) {
params.model = "models/" + string_split(params.hf_file, '/').back(); std::string cache_directory = get_cache_directory();
const bool success = create_directory_with_parents(cache_directory);
if (!success) {
throw std::runtime_error("failed to create cache directory: " + cache_directory);
}
params.model = cache_directory + string_split(params.hf_file, '/').back();
} }
} else if (!params.model_url.empty()) { } else if (!params.model_url.empty()) {
if (params.model.empty()) { if (params.model.empty()) {
@ -2517,6 +2522,31 @@ bool create_directory_with_parents(const std::string & path) {
#endif // _WIN32 #endif // _WIN32
} }
std::string get_cache_directory() {
std::string cache_directory = "";
if (getenv("LLAMA_CACHE")) {
cache_directory = std::getenv("LLAMA_CACHE");
if (cache_directory.back() != DIRECTORY_SEPARATOR) {
cache_directory += DIRECTORY_SEPARATOR;
}
} else {
#ifdef __linux__
if (std::getenv("XDG_CACHE_HOME")) {
cache_directory = std::getenv("XDG_CACHE_HOME");
} else {
cache_directory = std::getenv("HOME") + std::string("/.cache/");
}
#elif defined(__APPLE__)
cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
#elif defined(_WIN32)
cache_directory = std::getenv("APPDATA");
#endif // __linux__
cache_directory += "llama.cpp";
cache_directory += DIRECTORY_SEPARATOR;
}
return cache_directory;
}
void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector<float> & data) { void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector<float> & data) {
if (data.empty()) { if (data.empty()) {
fprintf(stream, "%s:\n", prop_name); fprintf(stream, "%s:\n", prop_name);

1
common/common.h
View file

@ -298,6 +298,7 @@ bool llama_should_add_bos_token(const llama_model * model);
// //
bool create_directory_with_parents(const std::string & path); bool create_directory_with_parents(const std::string & path);
std::string get_cache_directory();
void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector<float> & data); void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector<float> & data);
void dump_vector_int_yaml(FILE * stream, const char * prop_name, const std::vector<int> & data); void dump_vector_int_yaml(FILE * stream, const char * prop_name, const std::vector<int> & data);
void dump_string_yaml_multiline(FILE * stream, const char * prop_name, const char * data); void dump_string_yaml_multiline(FILE * stream, const char * prop_name, const char * data);

13
common/sampling.cpp
View file

@ -179,7 +179,7 @@ static llama_token llama_sampling_sample_impl(
struct llama_context * ctx_main, struct llama_context * ctx_main,
struct llama_context * ctx_cfg, struct llama_context * ctx_cfg,
const int idx, const int idx,
bool is_resampling) { // Add a parameter to indicate if we are resampling bool is_resampling) {
const llama_sampling_params & params = ctx_sampling->params; const llama_sampling_params & params = ctx_sampling->params;
const float temp = params.temp; const float temp = params.temp;
@ -188,8 +188,8 @@ static llama_token llama_sampling_sample_impl(
const float mirostat_eta = params.mirostat_eta; const float mirostat_eta = params.mirostat_eta;
std::vector<float> original_logits; std::vector<float> original_logits;
auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, !is_resampling, &original_logits); auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, /* apply_grammar= */ is_resampling, &original_logits);
if (!is_resampling) { if (ctx_sampling->grammar != NULL && !is_resampling) {
GGML_ASSERT(!original_logits.empty()); GGML_ASSERT(!original_logits.empty());
} }
llama_token id = 0; llama_token id = 0;
@ -252,7 +252,7 @@ static llama_token llama_sampling_sample_impl(
// Restore logits from the copy // Restore logits from the copy
std::copy(original_logits.begin(), original_logits.end(), logits); std::copy(original_logits.begin(), original_logits.end(), logits);
return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, true); // Pass true for is_resampling return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ true);
} }
} }
@ -285,7 +285,8 @@ static llama_token_data_array llama_sampling_prepare_impl(
// Get a pointer to the logits // Get a pointer to the logits
float * logits = llama_get_logits_ith(ctx_main, idx); float * logits = llama_get_logits_ith(ctx_main, idx);
if (apply_grammar && original_logits != NULL) { if (ctx_sampling->grammar != NULL && !apply_grammar) {
GGML_ASSERT(original_logits != NULL);
// Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this. // Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this.
*original_logits = {logits, logits + llama_n_vocab(llama_get_model(ctx_main))}; *original_logits = {logits, logits + llama_n_vocab(llama_get_model(ctx_main))};
} }
@ -342,7 +343,7 @@ llama_token llama_sampling_sample(
struct llama_context * ctx_cfg, struct llama_context * ctx_cfg,
const int idx) { const int idx) {
// Call the implementation function with is_resampling set to false by default // Call the implementation function with is_resampling set to false by default
return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, false); return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ false);
} }
llama_token_data_array llama_sampling_prepare( llama_token_data_array llama_sampling_prepare(

2
convert-hf-to-gguf-update.py
View file

@ -72,7 +72,7 @@ models = [
{"name": "mpt", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", }, {"name": "mpt", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
{"name": "starcoder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", }, {"name": "starcoder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
{"name": "gpt-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", }, {"name": "gpt-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },
{"name": "stablelm", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b", }, {"name": "stablelm2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b", },
{"name": "refact", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", }, {"name": "refact", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", },
{"name": "command-r", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereForAI/c4ai-command-r-v01", }, {"name": "command-r", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereForAI/c4ai-command-r-v01", },
{"name": "qwen2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", }, {"name": "qwen2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },

55
convert-hf-to-gguf.py
View file

@ -14,6 +14,7 @@ from pathlib import Path
from hashlib import sha256 from hashlib import sha256
from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Sequence, TypeVar, cast from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Sequence, TypeVar, cast
import math
import numpy as np import numpy as np
import torch import torch
@ -447,7 +448,7 @@ class Model:
# ref: https://huggingface.co/openai-community/gpt2 # ref: https://huggingface.co/openai-community/gpt2
res = "gpt-2" res = "gpt-2"
if chkhsh == "32d85c31273f8019248f2559fed492d929ea28b17e51d81d3bb36fff23ca72b3": if chkhsh == "32d85c31273f8019248f2559fed492d929ea28b17e51d81d3bb36fff23ca72b3":
# ref: https://huggingface.co/stabilityai/stablelm-2-1_6b # ref: https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b
res = "stablelm2" res = "stablelm2"
if chkhsh == "6221ad2852e85ce96f791f476e0b390cf9b474c9e3d1362f53a24a06dc8220ff": if chkhsh == "6221ad2852e85ce96f791f476e0b390cf9b474c9e3d1362f53a24a06dc8220ff":
# ref: https://huggingface.co/smallcloudai/Refact-1_6-base # ref: https://huggingface.co/smallcloudai/Refact-1_6-base
@ -1749,7 +1750,7 @@ class Phi3MiniModel(Model):
token_id = int(token_id) token_id = int(token_id)
token = foken_data["content"].encode("utf-8") token = foken_data["content"].encode("utf-8")
if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN: if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN:
assert(tokens[token_id] == token) assert tokens[token_id] == token
tokens[token_id] = token tokens[token_id] = token
scores[token_id] = -1000.0 scores[token_id] = -1000.0
toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
@ -1765,7 +1766,7 @@ class Phi3MiniModel(Model):
token_id = int(foken_data["id"]) token_id = int(foken_data["id"])
token = foken_data["content"].encode("utf-8") token = foken_data["content"].encode("utf-8")
if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN: if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN:
assert(tokens[token_id] == token) assert tokens[token_id] == token
tokens[token_id] = token tokens[token_id] = token
scores[token_id] = -1000.0 scores[token_id] = -1000.0
toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
@ -1784,23 +1785,59 @@ class Phi3MiniModel(Model):
def set_gguf_parameters(self): def set_gguf_parameters(self):
block_count = self.find_hparam(["num_hidden_layers", "n_layer"]) block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
rot_pct = 1.0
n_embd = self.find_hparam(["hidden_size", "n_embd"]) n_embd = self.find_hparam(["hidden_size", "n_embd"])
n_head = self.find_hparam(["num_attention_heads", "n_head"]) n_head = self.find_hparam(["num_attention_heads", "n_head"])
n_head_kv = self.find_hparam(["num_key_value_heads", "n_head_kv"])
rms_eps = self.find_hparam(["rms_norm_eps"]) rms_eps = self.find_hparam(["rms_norm_eps"])
max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"])
orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"])
rope_dims = n_embd // n_head
self.gguf_writer.add_name("Phi3") self.gguf_writer.add_name("Phi3")
self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"])) self.gguf_writer.add_context_length(max_pos_embds)
self.gguf_writer.add_rope_scaling_orig_ctx_len(orig_max_pos_embds)
self.gguf_writer.add_embedding_length(n_embd) self.gguf_writer.add_embedding_length(n_embd)
self.gguf_writer.add_feed_forward_length(8192) self.gguf_writer.add_feed_forward_length(self.find_hparam(["intermediate_size"]))
self.gguf_writer.add_block_count(block_count) self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_head_count(n_head) self.gguf_writer.add_head_count(n_head)
self.gguf_writer.add_head_count_kv(n_head) self.gguf_writer.add_head_count_kv(n_head_kv)
self.gguf_writer.add_layer_norm_rms_eps(rms_eps) self.gguf_writer.add_layer_norm_rms_eps(rms_eps)
self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head) self.gguf_writer.add_rope_dimension_count(rope_dims)
self.gguf_writer.add_rope_freq_base(self.find_hparam(["rope_theta"]))
self.gguf_writer.add_file_type(self.ftype) self.gguf_writer.add_file_type(self.ftype)
# write rope scaling for long context (128k) model
rope_scaling = self.find_hparam(['rope_scaling'], True)
if (rope_scaling is None):
return
scale = max_pos_embds / orig_max_pos_embds
rope_scaling_type = rope_scaling.get('type', '').lower()
if len(rope_scaling_type) == 0:
raise KeyError('Missing the required key rope_scaling.type')
if rope_scaling_type == 'su':
attn_factor = math.sqrt(1 + math.log(scale) / math.log(orig_max_pos_embds)) if scale > 1.0 else 1.0
elif rope_scaling_type == 'yarn':
attn_factor = 0.1 * math.log(scale) + 1.0 if scale > 1.0 else 1.0
else:
raise NotImplementedError(f'The rope scaling type {rope_scaling_type} is not supported yet')
self.gguf_writer.add_rope_scaling_attn_factors(attn_factor)
long_factors = rope_scaling.get('long_factor', None)
short_factors = rope_scaling.get('short_factor', None)
if long_factors is None or short_factors is None:
raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor')
if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2:
raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}')
self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG] + ".weight", np.array(long_factors, dtype=np.float32))
self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32))
@Model.register("PlamoForCausalLM") @Model.register("PlamoForCausalLM")
class PlamoModel(Model): class PlamoModel(Model):

4
examples/finetune/finetune.cpp
View file

@ -563,8 +563,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
// not capturing these, to silcence warnings // not capturing these, to silcence warnings
const int rope_mode = 0; const int rope_mode = 0;
return ggml_rope_custom(ctx, return ggml_rope_ext(ctx,
t, KQ_pos, n_rot, rope_mode, n_ctx, 0, t, KQ_pos, nullptr, n_rot, rope_mode, n_ctx, 0,
rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
); );
}; };

2
examples/main/README.md
View file

@ -325,3 +325,5 @@ These options provide extra functionality and customization when running the LLa
- `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. - `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance.
- `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains. - `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
- `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation. - `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
- `-hfr URL --hf-repo URL`: The url to the Hugging Face model repository. Used in conjunction with `--hf-file` or `-hff`. The model is downloaded and stored in the file provided by `-m` or `--model`. If `-m` is not provided, the model is auto-stored in the path specified by the `LLAMA_CACHE` environment variable or in an OS-specific local cache.

4
examples/main/main.cpp
View file

@ -708,7 +708,7 @@ int main(int argc, char ** argv) {
const llama_token id = llama_sampling_sample(ctx_sampling, ctx, ctx_guidance); const llama_token id = llama_sampling_sample(ctx_sampling, ctx, ctx_guidance);
llama_sampling_accept(ctx_sampling, ctx, id, true); llama_sampling_accept(ctx_sampling, ctx, id, /* apply_grammar= */ true);
LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str()); LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
@ -729,7 +729,7 @@ int main(int argc, char ** argv) {
// push the prompt in the sampling context in order to apply repetition penalties later // push the prompt in the sampling context in order to apply repetition penalties later
// for the prompt, we don't apply grammar rules // for the prompt, we don't apply grammar rules
llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], false); llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], /* apply_grammar= */ false);
++n_consumed; ++n_consumed;
if ((int) embd.size() >= params.n_batch) { if ((int) embd.size() >= params.n_batch) {

10
examples/server/tests/features/server.feature
View file

@ -37,8 +37,8 @@ Feature: llama.cpp server
Examples: Prompts Examples: Prompts
| prompt | n_predict | re_content | n_prompt | n_predicted | truncated | | prompt | n_predict | re_content | n_prompt | n_predicted | truncated |
| I believe the meaning of life is | 8 | (read\|going\|pretty)+ | 18 | 8 | not | | I believe the meaning of life is | 8 | (read\|going)+ | 18 | 8 | not |
| Write a joke about AI from a very long prompt which will not be truncated | 256 | (princesses\|everyone\|kids\|Anna\|forest)+ | 45 | 64 | not | | Write a joke about AI from a very long prompt which will not be truncated | 256 | (princesses\|everyone\|kids\|Anna\|forest)+ | 46 | 64 | not |
Scenario: Completion prompt truncated Scenario: Completion prompt truncated
Given a prompt: Given a prompt:
@ -67,8 +67,8 @@ Feature: llama.cpp server
Examples: Prompts Examples: Prompts
| model | system_prompt | user_prompt | max_tokens | re_content | n_prompt | n_predicted | enable_streaming | truncated | | model | system_prompt | user_prompt | max_tokens | re_content | n_prompt | n_predicted | enable_streaming | truncated |
| llama-2 | Book | What is the best book | 8 | (Here\|what)+ | 76 | 8 | disabled | not | | llama-2 | Book | What is the best book | 8 | (Here\|what)+ | 77 | 8 | disabled | not |
| codellama70b | You are a coding assistant. | Write the fibonacci function in c++. | 128 | (thanks\|happy\|bird\|fireplace)+ | -1 | 64 | enabled | | | codellama70b | You are a coding assistant. | Write the fibonacci function in c++. | 128 | (thanks\|happy\|bird\|Annabyear)+ | -1 | 64 | enabled | |
Scenario Outline: OAI Compatibility w/ response format Scenario Outline: OAI Compatibility w/ response format
@ -84,7 +84,7 @@ Feature: llama.cpp server
| response_format | n_predicted | re_content | | response_format | n_predicted | re_content |
| {"type": "json_object", "schema": {"const": "42"}} | 5 | "42" | | {"type": "json_object", "schema": {"const": "42"}} | 5 | "42" |
| {"type": "json_object", "schema": {"items": [{"type": "integer"}]}} | 10 | \[ -300 \] | | {"type": "json_object", "schema": {"items": [{"type": "integer"}]}} | 10 | \[ -300 \] |
| {"type": "json_object"} | 10 | \{ " Saragine. | | {"type": "json_object"} | 10 | \{ " Jacky. |
Scenario: Tokenize / Detokenize Scenario: Tokenize / Detokenize

4
examples/server/tests/features/slotsave.feature
View file

@ -26,7 +26,7 @@ Feature: llama.cpp server slot management
# Since we have cache, this should only process the last tokens # Since we have cache, this should only process the last tokens
Given a user prompt "What is the capital of Germany?" Given a user prompt "What is the capital of Germany?"
And a completion request with no api error And a completion request with no api error
Then 24 tokens are predicted matching (Thank|special|Lily) Then 24 tokens are predicted matching (Thank|special)
And 7 prompt tokens are processed And 7 prompt tokens are processed
# Loading the original cache into slot 0, # Loading the original cache into slot 0,
# we should only be processing 1 prompt token and get the same output # we should only be processing 1 prompt token and get the same output
@ -41,7 +41,7 @@ Feature: llama.cpp server slot management
Given a user prompt "What is the capital of Germany?" Given a user prompt "What is the capital of Germany?"
And using slot id 1 And using slot id 1
And a completion request with no api error And a completion request with no api error
Then 24 tokens are predicted matching (Thank|special|Lily) Then 24 tokens are predicted matching (Thank|special)
And 1 prompt tokens are processed And 1 prompt tokens are processed
Scenario: Erase Slot Scenario: Erase Slot

4
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -301,8 +301,8 @@ static struct ggml_tensor * llama_build_train_graphs(
// not capturing these, to silcence warnings // not capturing these, to silcence warnings
const int rope_mode = 0; const int rope_mode = 0;
return ggml_rope_custom( return ggml_rope_ext(
ctx, t, KQ_pos, n_rot, rope_mode, n_ctx, 0, rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f ctx, t, KQ_pos, nullptr, n_rot, rope_mode, n_ctx, 0, rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
); );
}; };

1247
ggml-cuda/mmq.cu
View file

File diff suppressed because it is too large Load diff

72
ggml-cuda/rope.cu
View file

@ -58,10 +58,10 @@ static __global__ void rope(
dst[i + 1] = x0*sin_theta + x1*cos_theta; dst[i + 1] = x0*sin_theta + x1*cos_theta;
} }
template<typename T, bool has_pos> template<typename T, bool has_pos, bool has_freq_facs>
static __global__ void rope_neox( static __global__ void rope_neox(
const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows, const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims, const float * freq_factors
) { ) {
const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y); const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
@ -88,7 +88,9 @@ static __global__ void rope_neox(
float cur_rot = inv_ndims * ic - ib; float cur_rot = inv_ndims * ic - ib;
const int p = has_pos ? pos[i2] : 0; const int p = has_pos ? pos[i2] : 0;
const float theta_base = p*freq_scale*powf(theta_scale, col/2.0f); const float freq_factor = has_freq_facs ? freq_factors[ic/2] : 1.0f;
const float theta_base = p*freq_scale*powf(theta_scale, col/2.0f)/freq_factor;
float cos_theta, sin_theta; float cos_theta, sin_theta;
rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta); rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
@ -164,7 +166,7 @@ static void rope_cuda(
template<typename T> template<typename T>
static void rope_neox_cuda( static void rope_neox_cuda(
const T * x, T * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, const T * x, T * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream
) { ) {
GGML_ASSERT(ncols % 2 == 0); GGML_ASSERT(ncols % 2 == 0);
const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1); const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
@ -175,15 +177,29 @@ static void rope_neox_cuda(
const float inv_ndims = -1.0f / n_dims; const float inv_ndims = -1.0f / n_dims;
if (pos == nullptr) { if (pos == nullptr) {
rope_neox<T, false><<<block_nums, block_dims, 0, stream>>>( if (freq_factors == nullptr) {
x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims, rope_neox<T, false, false><<<block_nums, block_dims, 0, stream>>>(
theta_scale, inv_ndims x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
); theta_scale, inv_ndims, freq_factors
);
} else {
rope_neox<T, false, true><<<block_nums, block_dims, 0, stream>>>(
x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
theta_scale, inv_ndims, freq_factors
);
}
} else { } else {
rope_neox<T, true><<<block_nums, block_dims, 0, stream>>>( if (freq_factors == nullptr) {
x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims, rope_neox<T, true, false><<<block_nums, block_dims, 0, stream>>>(
theta_scale, inv_ndims x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
); theta_scale, inv_ndims, freq_factors
);
} else {
rope_neox<T, true, true><<<block_nums, block_dims, 0, stream>>>(
x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
theta_scale, inv_ndims, freq_factors
);
}
} }
} }
@ -214,24 +230,27 @@ static void rope_cuda_f32(
static void rope_neox_cuda_f16( static void rope_neox_cuda_f16(
const half * x, half * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, const half * x, half * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream) { float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream) {
rope_neox_cuda<half>(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream); rope_neox_cuda<half>(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
} }
static void rope_neox_cuda_f32( static void rope_neox_cuda_f32(
const float * x, float * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, const float * x, float * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream
) { ) {
rope_neox_cuda<float>(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream); rope_neox_cuda<float>(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
} }
void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0]; const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1]; const ggml_tensor * src1 = dst->src[1];
const ggml_tensor * src2 = dst->src[2];
const float * src0_d = (const float *)src0->data; const float * src0_d = (const float *)src0->data;
const float * src1_d = (const float *)src1->data; const float * src1_d = (const float *)src1->data;
float * dst_d = (float *)dst->data; float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream(); cudaStream_t stream = ctx.stream();
@ -241,7 +260,6 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const int64_t ne00 = src0->ne[0]; const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1]; const int64_t ne01 = src0->ne[1];
const int64_t ne2 = dst->ne[2];
const int64_t nrows = ggml_nrows(src0); const int64_t nrows = ggml_nrows(src0);
//const int n_past = ((int32_t *) dst->op_params)[0]; //const int n_past = ((int32_t *) dst->op_params)[0];
@ -259,16 +277,22 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
const float * freq_factors = nullptr;
const int32_t * pos = nullptr; const int32_t * pos = nullptr;
if ((mode & 1) == 0) {
GGML_ASSERT(src1->type == GGML_TYPE_I32);
GGML_ASSERT(src1->ne[0] == ne2);
pos = (const int32_t *) src1_d;
}
const bool is_neox = mode & 2; const bool is_neox = mode & 2;
const bool is_glm = mode & 4; const bool is_glm = mode & 4;
if (is_neox) {
pos = (const int32_t *) src1_d;
if (src2 != nullptr) {
freq_factors = (const float *) src2->data;
}
} else {
GGML_ASSERT(src2 == nullptr && "TODO: freq_factors not implemented for !is_neox");
}
rope_corr_dims corr_dims; rope_corr_dims corr_dims;
ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims.v); ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims.v);
@ -280,12 +304,12 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
if (src0->type == GGML_TYPE_F32) { if (src0->type == GGML_TYPE_F32) {
rope_neox_cuda_f32( rope_neox_cuda_f32(
(const float *)src0_d, (float *)dst_d, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor, (const float *)src0_d, (float *)dst_d, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
attn_factor, corr_dims, stream attn_factor, corr_dims, freq_factors, stream
); );
} else if (src0->type == GGML_TYPE_F16) { } else if (src0->type == GGML_TYPE_F16) {
rope_neox_cuda_f16( rope_neox_cuda_f16(
(const half *)src0_d, (half *)dst_d, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor, (const half *)src0_d, (half *)dst_d, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
attn_factor, corr_dims, stream attn_factor, corr_dims, freq_factors, stream
); );
} else { } else {
GGML_ASSERT(false); GGML_ASSERT(false);

4
ggml-kompute.cpp
View file

@ -1677,6 +1677,10 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
} break; } break;
case GGML_OP_ROPE: case GGML_OP_ROPE:
{ {
#pragma message("TODO: implement phi3 frequency factors support")
#pragma message(" https://github.com/ggerganov/llama.cpp/pull/7225")
GGML_ASSERT(dst->src[2] == nullptr && "phi3 frequency factors not implemented yet");
GGML_ASSERT(ne10 == ne02); GGML_ASSERT(ne10 == ne02);
GGML_ASSERT(src0t == dstt); GGML_ASSERT(src0t == dstt);
// const int n_past = ((int32_t *) dst->op_params)[0]; // const int n_past = ((int32_t *) dst->op_params)[0];

121
ggml-metal.m
View file

@ -927,22 +927,32 @@ static enum ggml_status ggml_metal_graph_compute(
const int64_t ne10 = src1 ? src1->ne[0] : 0; const int64_t ne10 = src1 ? src1->ne[0] : 0;
const int64_t ne11 = src1 ? src1->ne[1] : 0; const int64_t ne11 = src1 ? src1->ne[1] : 0;
const int64_t ne12 = src1 ? src1->ne[2] : 0; const int64_t ne12 = src1 ? src1->ne[2] : 0;
const int64_t ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13); const int64_t ne13 = src1 ? src1->ne[3] : 0;
const uint64_t nb10 = src1 ? src1->nb[0] : 0; const uint64_t nb10 = src1 ? src1->nb[0] : 0;
const uint64_t nb11 = src1 ? src1->nb[1] : 0; const uint64_t nb11 = src1 ? src1->nb[1] : 0;
const uint64_t nb12 = src1 ? src1->nb[2] : 0; const uint64_t nb12 = src1 ? src1->nb[2] : 0;
const uint64_t nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13); const uint64_t nb13 = src1 ? src1->nb[3] : 0;
const int64_t ne0 = dst ? dst->ne[0] : 0; const int64_t ne20 = src2 ? src2->ne[0] : 0;
const int64_t ne1 = dst ? dst->ne[1] : 0; const int64_t ne21 = src2 ? src2->ne[1] : 0;
const int64_t ne2 = dst ? dst->ne[2] : 0; const int64_t ne22 = src2 ? src2->ne[2] : 0; GGML_UNUSED(ne22);
const int64_t ne3 = dst ? dst->ne[3] : 0; const int64_t ne23 = src2 ? src2->ne[3] : 0; GGML_UNUSED(ne23);
const uint64_t nb0 = dst ? dst->nb[0] : 0; const uint64_t nb20 = src2 ? src2->nb[0] : 0; GGML_UNUSED(nb20);
const uint64_t nb1 = dst ? dst->nb[1] : 0; const uint64_t nb21 = src2 ? src2->nb[1] : 0;
const uint64_t nb2 = dst ? dst->nb[2] : 0; const uint64_t nb22 = src2 ? src2->nb[2] : 0;
const uint64_t nb3 = dst ? dst->nb[3] : 0; const uint64_t nb23 = src2 ? src2->nb[3] : 0;
const int64_t ne0 = dst ? dst->ne[0] : 0;
const int64_t ne1 = dst ? dst->ne[1] : 0;
const int64_t ne2 = dst ? dst->ne[2] : 0;
const int64_t ne3 = dst ? dst->ne[3] : 0;
const uint64_t nb0 = dst ? dst->nb[0] : 0;
const uint64_t nb1 = dst ? dst->nb[1] : 0;
const uint64_t nb2 = dst ? dst->nb[2] : 0;
const uint64_t nb3 = dst ? dst->nb[3] : 0;
const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT; const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT; const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
@ -1785,16 +1795,6 @@ static enum ggml_status ggml_metal_graph_compute(
const int n_as = src0->ne[2]; const int n_as = src0->ne[2];
// src2 = ids // src2 = ids
const int64_t ne20 = src2->ne[0];
const int64_t ne21 = src2->ne[1];
const int64_t ne22 = src2->ne[2]; GGML_UNUSED(ne22);
const int64_t ne23 = src2->ne[3]; GGML_UNUSED(ne23);
const uint64_t nb20 = src2->nb[0]; GGML_UNUSED(nb20);
const uint64_t nb21 = src2->nb[1];
const uint64_t nb22 = src2->nb[2]; GGML_UNUSED(nb22);
const uint64_t nb23 = src2->nb[3]; GGML_UNUSED(nb23);
const enum ggml_type src2t = src2->type; GGML_UNUSED(src2t); const enum ggml_type src2t = src2->type; GGML_UNUSED(src2t);
GGML_ASSERT(src2t == GGML_TYPE_I32); GGML_ASSERT(src2t == GGML_TYPE_I32);
@ -2244,7 +2244,13 @@ static enum ggml_status ggml_metal_graph_compute(
// skip 3, n_ctx, used in GLM RoPE, unimplemented in metal // skip 3, n_ctx, used in GLM RoPE, unimplemented in metal
const int n_orig_ctx = ((int32_t *) dst->op_params)[4]; const int n_orig_ctx = ((int32_t *) dst->op_params)[4];
float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; float freq_base;
float freq_scale;
float ext_factor;
float attn_factor;
float beta_fast;
float beta_slow;
memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float));
memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float));
memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
@ -2252,6 +2258,15 @@ static enum ggml_status ggml_metal_graph_compute(
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
const bool is_neox = mode & 2;
const bool is_glm = mode & 4;
GGML_ASSERT(!is_glm && "GLM RoPE not implemented in Metal");
if (!is_neox) {
GGML_ASSERT(id_src2 == nil && "TODO: freq_factors not implemented for !is_neox");
}
id<MTLComputePipelineState> pipeline = nil; id<MTLComputePipelineState> pipeline = nil;
switch (src0->type) { switch (src0->type) {
@ -2263,33 +2278,38 @@ static enum ggml_status ggml_metal_graph_compute(
[encoder setComputePipelineState:pipeline]; [encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2]; if (id_src2 != nil) {
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:3]; [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:4]; } else {
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:5]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:2];
[encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:6]; }
[encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:7]; [encoder setBuffer:id_dst offset:offs_dst atIndex:3];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:8]; [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:4];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:9]; [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:5];
[encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:10]; [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:6];
[encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:11]; [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:7];
[encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:12]; [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:8];
[encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:13]; [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:14]; [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:10];
[encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:15]; [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:11];
[encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:16]; [encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:12];
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:17]; [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:13];
[encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:18]; [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:14];
[encoder setBytes:&n_past length:sizeof( int) atIndex:19]; [encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:15];
[encoder setBytes:&n_dims length:sizeof( int) atIndex:20]; [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&mode length:sizeof( int) atIndex:21]; [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:17];
[encoder setBytes:&n_orig_ctx length:sizeof( int) atIndex:22]; [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:18];
[encoder setBytes:&freq_base length:sizeof( float) atIndex:23]; [encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:19];
[encoder setBytes:&freq_scale length:sizeof( float) atIndex:24]; [encoder setBytes:&n_past length:sizeof( int) atIndex:20];
[encoder setBytes:&ext_factor length:sizeof( float) atIndex:25]; [encoder setBytes:&n_dims length:sizeof( int) atIndex:21];
[encoder setBytes:&attn_factor length:sizeof( float) atIndex:26]; [encoder setBytes:&mode length:sizeof( int) atIndex:22];
[encoder setBytes:&beta_fast length:sizeof( float) atIndex:27]; [encoder setBytes:&n_orig_ctx length:sizeof( int) atIndex:23];
[encoder setBytes:&beta_slow length:sizeof( float) atIndex:28]; [encoder setBytes:&freq_base length:sizeof( float) atIndex:24];
[encoder setBytes:&freq_scale length:sizeof( float) atIndex:25];
[encoder setBytes:&ext_factor length:sizeof( float) atIndex:26];
[encoder setBytes:&attn_factor length:sizeof( float) atIndex:27];
[encoder setBytes:&beta_fast length:sizeof( float) atIndex:28];
[encoder setBytes:&beta_slow length:sizeof( float) atIndex:29];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break; } break;
@ -2535,11 +2555,6 @@ static enum ggml_status ggml_metal_graph_compute(
GGML_ASSERT(!src3 || src3->ne[1] >= GGML_PAD(src0->ne[1], 8) && GGML_ASSERT(!src3 || src3->ne[1] >= GGML_PAD(src0->ne[1], 8) &&
"the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big"); "the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big");
const uint64_t nb20 = src2 ? src2->nb[0] : 0; GGML_UNUSED(nb20);
const uint64_t nb21 = src2 ? src2->nb[1] : 0;
const uint64_t nb22 = src2 ? src2->nb[2] : 0;
const uint64_t nb23 = src2 ? src2->nb[3] : 0;
const int64_t ne30 = src3 ? src3->ne[0] : 0; GGML_UNUSED(ne30); const int64_t ne30 = src3 ? src3->ne[0] : 0; GGML_UNUSED(ne30);
//const int64_t ne31 = src3 ? src3->ne[1] : 0; //const int64_t ne31 = src3 ? src3->ne[1] : 0;
const int64_t ne32 = src3 ? src3->ne[2] : 0; GGML_UNUSED(ne32); const int64_t ne32 = src3 ? src3->ne[2] : 0; GGML_UNUSED(ne32);

33
ggml-metal.metal
View file

@ -1640,6 +1640,7 @@ static void rope_yarn_corr_dims(
typedef void (rope_t)( typedef void (rope_t)(
device const void * src0, device const void * src0,
device const int32_t * src1, device const int32_t * src1,
device const float * src2,
device float * dst, device float * dst,
constant int64_t & ne00, constant int64_t & ne00,
constant int64_t & ne01, constant int64_t & ne01,
@ -1675,6 +1676,7 @@ template<typename T>
kernel void kernel_rope( kernel void kernel_rope(
device const void * src0, device const void * src0,
device const int32_t * src1, device const int32_t * src1,
device const float * src2,
device float * dst, device float * dst,
constant int64_t & ne00, constant int64_t & ne00,
constant int64_t & ne01, constant int64_t & ne01,
@ -1744,8 +1746,10 @@ kernel void kernel_rope(
// simplified from `(ib * n_dims + ic) * inv_ndims` // simplified from `(ib * n_dims + ic) * inv_ndims`
const float cur_rot = inv_ndims*ic - ib; const float cur_rot = inv_ndims*ic - ib;
const float freq_factor = src2 != src0 ? src2[ic/2] : 1.0f;
const float theta = theta_0 * pow(freq_base, cur_rot) / freq_factor;
const float theta = theta_0 * pow(freq_base, cur_rot);
float cos_theta, sin_theta; float cos_theta, sin_theta;
rope_yarn(theta, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta); rope_yarn(theta, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
@ -2204,11 +2208,7 @@ kernel void kernel_flash_attn_ext_f16(
// pointer to the mask // pointer to the mask
device const half * mp = (device const half *) (mask + iq1*nb31); device const half * mp = (device const half *) (mask + iq1*nb31);
// prepare diagonal scale matrix float slope = 1.0f;
simdgroup_float8x8 mscale(scale);
// prepare diagonal slope matrix
simdgroup_float8x8 mslope(1.0f);
// ALiBi // ALiBi
if (max_bias > 0.0f) { if (max_bias > 0.0f) {
@ -2217,7 +2217,7 @@ kernel void kernel_flash_attn_ext_f16(
const float base = h < n_head_log2 ? m0 : m1; const float base = h < n_head_log2 ? m0 : m1;
const int exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; const int exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
mslope = simdgroup_float8x8(pow(base, exph)); slope = pow(base, exph);
} }
// loop over the KV cache // loop over the KV cache
@ -2242,18 +2242,20 @@ kernel void kernel_flash_attn_ext_f16(
simdgroup_multiply_accumulate(mqk, mq[i], mk, mqk); simdgroup_multiply_accumulate(mqk, mq[i], mk, mqk);
} }
simdgroup_store(mqk, ss + 8*cc, TF, 0, false);
const short tx = tiisg%4;
const short ty = tiisg/4;
if (mask != q) { if (mask != q) {
// mqk = mqk*scale + mask*slope // mqk = mqk*scale + mask*slope
simdgroup_half8x8 mm; ss[8*cc + ty*TF + 2*tx + 0] = scale*ss[8*cc + ty*TF + 2*tx + 0] + slope*mp[ic + 8*cc + ty*nb31/sizeof(half) + 2*tx + 0];
simdgroup_load(mm, mp + ic + 8*cc, nb31/sizeof(half), 0, false); ss[8*cc + ty*TF + 2*tx + 1] = scale*ss[8*cc + ty*TF + 2*tx + 1] + slope*mp[ic + 8*cc + ty*nb31/sizeof(half) + 2*tx + 1];
simdgroup_multiply(mm, mslope, mm);
simdgroup_multiply_accumulate(mqk, mqk, mscale, mm);
} else { } else {
// mqk = mqk*scale // mqk = mqk*scale
simdgroup_multiply(mqk, mscale, mqk); ss[8*cc + ty*TF + 2*tx + 0] *= scale;
ss[8*cc + ty*TF + 2*tx + 1] *= scale;
} }
simdgroup_store(mqk, ss + 8*cc, TF, 0, false);
} }
} }
@ -2816,8 +2818,7 @@ kernel void kernel_cpy_f32_f16(
for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) { for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00); device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
// TODO: is there a better way to handle -INFINITY? dst_data[i00] = src[0];
dst_data[i00] = src[0] == -INFINITY ? -MAXHALF : src[0];
} }
} }

3
ggml-sycl.cpp
View file

@ -14454,6 +14454,9 @@ inline void ggml_sycl_op_rope(const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd, ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd, const float *src1_dd, float *dst_dd,
const dpct::queue_ptr &main_stream) { const dpct::queue_ptr &main_stream) {
#pragma message("TODO: implement phi3 frequency factors support")
#pragma message(" https://github.com/ggerganov/llama.cpp/pull/7225")
GGML_ASSERT(dst->src[2] == nullptr && "phi3 frequency factors not implemented yet");
GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);

4
ggml-vulkan.cpp
View file

@ -4238,6 +4238,10 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context * subctx,
} }
static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
#pragma message("TODO: implement phi3 frequency factors support")
#pragma message(" https://github.com/ggerganov/llama.cpp/pull/7225")
GGML_ASSERT(dst->src[2] == nullptr && "phi3 frequency factors not implemented yet");
const int n_dims = ((int32_t *) dst->op_params)[1]; const int n_dims = ((int32_t *) dst->op_params)[1];
const int mode = ((int32_t *) dst->op_params)[2]; const int mode = ((int32_t *) dst->op_params)[2];
// const int n_ctx = ((int32_t *) dst->op_params)[3]; // const int n_ctx = ((int32_t *) dst->op_params)[3];

88
ggml.c
View file

@ -6234,6 +6234,7 @@ static struct ggml_tensor * ggml_rope_impl(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
struct ggml_tensor * b, struct ggml_tensor * b,
struct ggml_tensor * c,
int n_dims, int n_dims,
int mode, int mode,
int n_ctx, int n_ctx,
@ -6251,6 +6252,11 @@ static struct ggml_tensor * ggml_rope_impl(
GGML_ASSERT(b->type == GGML_TYPE_I32); GGML_ASSERT(b->type == GGML_TYPE_I32);
GGML_ASSERT(a->ne[2] == b->ne[0]); GGML_ASSERT(a->ne[2] == b->ne[0]);
if (c) {
GGML_ASSERT(c->type == GGML_TYPE_F32);
GGML_ASSERT(c->ne[0] >= n_dims / 2);
}
bool is_node = false; bool is_node = false;
if (a->grad) { if (a->grad) {
@ -6274,6 +6280,7 @@ static struct ggml_tensor * ggml_rope_impl(
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = a; result->src[0] = a;
result->src[1] = b; result->src[1] = b;
result->src[2] = c;
return result; return result;
} }
@ -6286,7 +6293,7 @@ struct ggml_tensor * ggml_rope(
int mode, int mode,
int n_ctx) { int n_ctx) {
return ggml_rope_impl( return ggml_rope_impl(
ctx, a, b, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, false, false ctx, a, b, NULL, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, false, false
); );
} }
@ -6298,7 +6305,49 @@ struct ggml_tensor * ggml_rope_inplace(
int mode, int mode,
int n_ctx) { int n_ctx) {
return ggml_rope_impl( return ggml_rope_impl(
ctx, a, b, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, false, true ctx, a, b, NULL, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, false, true
);
}
struct ggml_tensor * ggml_rope_ext(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
int n_dims,
int mode,
int n_ctx,
int n_orig_ctx,
float freq_base,
float freq_scale,
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow) {
return ggml_rope_impl(
ctx, a, b, c, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow, 0.0f, false, false
);
}
struct ggml_tensor * ggml_rope_ext_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
int n_dims,
int mode,
int n_ctx,
int n_orig_ctx,
float freq_base,
float freq_scale,
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow) {
return ggml_rope_impl(
ctx, a, b, c, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow, 0.0f, false, true
); );
} }
@ -6317,7 +6366,7 @@ struct ggml_tensor * ggml_rope_custom(
float beta_fast, float beta_fast,
float beta_slow) { float beta_slow) {
return ggml_rope_impl( return ggml_rope_impl(
ctx, a, b, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale, ctx, a, b, NULL, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow, 0.0f, false, false ext_factor, attn_factor, beta_fast, beta_slow, 0.0f, false, false
); );
} }
@ -6337,27 +6386,18 @@ struct ggml_tensor * ggml_rope_custom_inplace(
float beta_fast, float beta_fast,
float beta_slow) { float beta_slow) {
return ggml_rope_impl( return ggml_rope_impl(
ctx, a, b, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale, ctx, a, b, NULL, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow, 0.0f, false, true ext_factor, attn_factor, beta_fast, beta_slow, 0.0f, false, true
); );
} }
struct ggml_tensor * ggml_rope_xpos_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int n_dims,
float base,
bool down) {
return ggml_rope_impl(ctx, a, b, n_dims, 0, 0, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, base, down, true);
}
// ggml_rope_back // ggml_rope_back
struct ggml_tensor * ggml_rope_back( struct ggml_tensor * ggml_rope_back(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
struct ggml_tensor * b, struct ggml_tensor * b,
struct ggml_tensor * c,
int n_dims, int n_dims,
int mode, int mode,
int n_ctx, int n_ctx,
@ -6373,6 +6413,7 @@ struct ggml_tensor * ggml_rope_back(
GGML_ASSERT(ggml_is_vector(b)); GGML_ASSERT(ggml_is_vector(b));
GGML_ASSERT(b->type == GGML_TYPE_I32); GGML_ASSERT(b->type == GGML_TYPE_I32);
GGML_ASSERT(a->ne[2] == b->ne[0]); GGML_ASSERT(a->ne[2] == b->ne[0]);
GGML_ASSERT(c == NULL && "freq factors not implemented yet");
GGML_ASSERT((mode & 4) == 0 && "ggml_rope_back() for ChatGLM not implemented yet"); GGML_ASSERT((mode & 4) == 0 && "ggml_rope_back() for ChatGLM not implemented yet");
@ -14307,6 +14348,7 @@ static void ggml_compute_forward_rope_f32(
const struct ggml_tensor * src0 = dst->src[0]; const struct ggml_tensor * src0 = dst->src[0];
const struct ggml_tensor * src1 = dst->src[1]; const struct ggml_tensor * src1 = dst->src[1];
const struct ggml_tensor * src2 = dst->src[2];
if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) { if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
return; return;
@ -14366,6 +14408,17 @@ static void ggml_compute_forward_rope_f32(
const bool is_neox = mode & 2; const bool is_neox = mode & 2;
const bool is_glm = mode & 4; const bool is_glm = mode & 4;
const float * freq_factors = NULL;
if (is_neox) {
if (src2 != NULL) {
GGML_ASSERT(src2->type == GGML_TYPE_F32);
GGML_ASSERT(src2->ne[0] >= n_dims / 2);
freq_factors = (const float *) src2->data;
}
} else {
GGML_ASSERT(src2 == NULL && "TODO: freq_factors not implemented for mode 1");
}
// backward process uses inverse rotation by cos and sin. // backward process uses inverse rotation by cos and sin.
// cos and sin build a rotation matrix, where the inverse is the transpose. // cos and sin build a rotation matrix, where the inverse is the transpose.
// this essentially just switches the sign of sin. // this essentially just switches the sign of sin.
@ -14442,10 +14495,11 @@ static void ggml_compute_forward_rope_f32(
// simplified from `(ib * n_dims + ic) * inv_ndims` // simplified from `(ib * n_dims + ic) * inv_ndims`
float cur_rot = inv_ndims * ic - ib; float cur_rot = inv_ndims * ic - ib;
float freq_factor = freq_factors ? freq_factors[ic/2] : 1.0f;
float cos_theta, sin_theta; float cos_theta, sin_theta;
rope_yarn( rope_yarn(
theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, theta_base/freq_factor, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor,
&cos_theta, &sin_theta &cos_theta, &sin_theta
); );
sin_theta *= sin_sign; sin_theta *= sin_sign;
@ -18390,6 +18444,7 @@ static struct ggml_tensor * ggml_sub_or_set(struct ggml_context * ctx, struct gg
static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, struct ggml_hash_set zero_table) { static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, struct ggml_hash_set zero_table) {
struct ggml_tensor * src0 = tensor->src[0]; struct ggml_tensor * src0 = tensor->src[0];
struct ggml_tensor * src1 = tensor->src[1]; struct ggml_tensor * src1 = tensor->src[1];
struct ggml_tensor * src2 = tensor->src[2];
switch (tensor->op) { switch (tensor->op) {
case GGML_OP_DUP: case GGML_OP_DUP:
@ -18921,6 +18976,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
ggml_rope_back(ctx, ggml_rope_back(ctx,
tensor->grad, tensor->grad,
src1, src1,
src2,
n_dims, n_dims,
mode, mode,
n_ctx, n_ctx,
@ -18960,6 +19016,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
ggml_rope_impl(ctx, ggml_rope_impl(ctx,
tensor->grad, tensor->grad,
src1, src1,
src2,
n_dims, n_dims,
mode, mode,
n_ctx, n_ctx,
@ -19041,7 +19098,6 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
masked); masked);
} }
struct ggml_tensor * src2 = tensor->src[2];
const int64_t elem_q = ggml_nelements(src0); const int64_t elem_q = ggml_nelements(src0);
const int64_t elem_k = ggml_nelements(src1); const int64_t elem_k = ggml_nelements(src1);
const int64_t elem_v = ggml_nelements(src2); const int64_t elem_v = ggml_nelements(src2);

49
ggml.h
View file

@ -1472,6 +1472,7 @@ extern "C" {
// if mode & 4 == 1, ChatGLM style // if mode & 4 == 1, ChatGLM style
// //
// b is an int32 vector with size a->ne[2], it contains the positions // b is an int32 vector with size a->ne[2], it contains the positions
// c is freq factors (e.g. phi3-128k), (optional)
GGML_API struct ggml_tensor * ggml_rope( GGML_API struct ggml_tensor * ggml_rope(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
@ -1490,10 +1491,11 @@ extern "C" {
int n_ctx); int n_ctx);
// custom RoPE // custom RoPE
GGML_API struct ggml_tensor * ggml_rope_custom( GGML_API struct ggml_tensor * ggml_rope_ext(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
struct ggml_tensor * b, struct ggml_tensor * b,
struct ggml_tensor * c,
int n_dims, int n_dims,
int mode, int mode,
int n_ctx, int n_ctx,
@ -1506,7 +1508,23 @@ extern "C" {
float beta_slow); float beta_slow);
// in-place, returns view(a) // in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_rope_custom_inplace( GGML_API struct ggml_tensor * ggml_rope_ext_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
int n_dims,
int mode,
int n_ctx,
int n_orig_ctx,
float freq_base,
float freq_scale,
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow);
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
struct ggml_tensor * b, struct ggml_tensor * b,
@ -1519,20 +1537,28 @@ extern "C" {
float ext_factor, float ext_factor,
float attn_factor, float attn_factor,
float beta_fast, float beta_fast,
float beta_slow); float beta_slow),
"use ggml_rope_ext instead");
// compute correction dims for YaRN RoPE scaling GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
GGML_CALL void ggml_rope_yarn_corr_dims(
int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]);
// xPos RoPE, in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_rope_xpos_inplace(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
struct ggml_tensor * b, struct ggml_tensor * b,
int n_dims, int n_dims,
float base, int mode,
bool down); int n_ctx,
int n_orig_ctx,
float freq_base,
float freq_scale,
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow),
"use ggml_rope_ext_inplace instead");
// compute correction dims for YaRN RoPE scaling
GGML_CALL void ggml_rope_yarn_corr_dims(
int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]);
// rotary position embedding backward, i.e compute dx from dy // rotary position embedding backward, i.e compute dx from dy
// a - dy // a - dy
@ -1540,6 +1566,7 @@ extern "C" {
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * a, struct ggml_tensor * a,
struct ggml_tensor * b, struct ggml_tensor * b,
struct ggml_tensor * c,
int n_dims, int n_dims,
int mode, int mode,
int n_ctx, int n_ctx,

17
gguf-py/gguf/constants.py
View file

@ -57,12 +57,13 @@ class Keys:
CAUSAL = "{arch}.attention.causal" CAUSAL = "{arch}.attention.causal"
class Rope: class Rope:
DIMENSION_COUNT = "{arch}.rope.dimension_count" DIMENSION_COUNT = "{arch}.rope.dimension_count"
FREQ_BASE = "{arch}.rope.freq_base" FREQ_BASE = "{arch}.rope.freq_base"
SCALING_TYPE = "{arch}.rope.scaling.type" SCALING_TYPE = "{arch}.rope.scaling.type"
SCALING_FACTOR = "{arch}.rope.scaling.factor" SCALING_FACTOR = "{arch}.rope.scaling.factor"
SCALING_ORIG_CTX_LEN = "{arch}.rope.scaling.original_context_length" SCALING_ATTN_FACTOR = "{arch}.rope.scaling.attn_factor"
SCALING_FINETUNED = "{arch}.rope.scaling.finetuned" SCALING_ORIG_CTX_LEN = "{arch}.rope.scaling.original_context_length"
SCALING_FINETUNED = "{arch}.rope.scaling.finetuned"
class SSM: class SSM:
CONV_KERNEL = "{arch}.ssm.conv_kernel" CONV_KERNEL = "{arch}.ssm.conv_kernel"
@ -148,6 +149,8 @@ class MODEL_TENSOR(IntEnum):
OUTPUT = auto() OUTPUT = auto()
OUTPUT_NORM = auto() OUTPUT_NORM = auto()
ROPE_FREQS = auto() ROPE_FREQS = auto()
ROPE_FACTORS_LONG = auto()
ROPE_FACTORS_SHORT = auto()
ATTN_Q = auto() ATTN_Q = auto()
ATTN_K = auto() ATTN_K = auto()
ATTN_V = auto() ATTN_V = auto()
@ -225,6 +228,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.OUTPUT_NORM: "output_norm", MODEL_TENSOR.OUTPUT_NORM: "output_norm",
MODEL_TENSOR.OUTPUT: "output", MODEL_TENSOR.OUTPUT: "output",
MODEL_TENSOR.ROPE_FREQS: "rope_freqs", MODEL_TENSOR.ROPE_FREQS: "rope_freqs",
MODEL_TENSOR.ROPE_FACTORS_LONG: "rope_factors_long",
MODEL_TENSOR.ROPE_FACTORS_SHORT: "rope_factors_short",
MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm", MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm",
MODEL_TENSOR.ATTN_NORM_2: "blk.{bid}.attn_norm_2", MODEL_TENSOR.ATTN_NORM_2: "blk.{bid}.attn_norm_2",
MODEL_TENSOR.ATTN_QKV: "blk.{bid}.attn_qkv", MODEL_TENSOR.ATTN_QKV: "blk.{bid}.attn_qkv",

3
gguf-py/gguf/gguf_writer.py
View file

@ -433,6 +433,9 @@ class GGUFWriter:
def add_rope_scaling_factor(self, value: float) -> None: def add_rope_scaling_factor(self, value: float) -> None:
self.add_float32(Keys.Rope.SCALING_FACTOR.format(arch=self.arch), value) self.add_float32(Keys.Rope.SCALING_FACTOR.format(arch=self.arch), value)
def add_rope_scaling_attn_factors(self, value: Sequence[float]) -> None:
self.add_float32(Keys.Rope.SCALING_ATTN_FACTOR.format(arch=self.arch), value)
def add_rope_scaling_orig_ctx_len(self, value: int) -> None: def add_rope_scaling_orig_ctx_len(self, value: int) -> None:
self.add_uint32(Keys.Rope.SCALING_ORIG_CTX_LEN.format(arch=self.arch), value) self.add_uint32(Keys.Rope.SCALING_ORIG_CTX_LEN.format(arch=self.arch), value)

288
llama.cpp
View file

@ -326,6 +326,7 @@ enum llm_kv {
LLM_KV_ROPE_SCALE_LINEAR, LLM_KV_ROPE_SCALE_LINEAR,
LLM_KV_ROPE_SCALING_TYPE, LLM_KV_ROPE_SCALING_TYPE,
LLM_KV_ROPE_SCALING_FACTOR, LLM_KV_ROPE_SCALING_FACTOR,
LLM_KV_ROPE_SCALING_ATTN_FACTOR,
LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,
LLM_KV_ROPE_SCALING_FINETUNED, LLM_KV_ROPE_SCALING_FINETUNED,
@ -403,6 +404,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" }, { LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" },
{ LLM_KV_ROPE_SCALING_TYPE, "%s.rope.scaling.type" }, { LLM_KV_ROPE_SCALING_TYPE, "%s.rope.scaling.type" },
{ LLM_KV_ROPE_SCALING_FACTOR, "%s.rope.scaling.factor" }, { LLM_KV_ROPE_SCALING_FACTOR, "%s.rope.scaling.factor" },
{ LLM_KV_ROPE_SCALING_ATTN_FACTOR, "%s.rope.scaling.attn_factor" },
{ LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" }, { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" },
{ LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" }, { LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" },
@ -458,6 +460,8 @@ enum llm_tensor {
LLM_TENSOR_OUTPUT, LLM_TENSOR_OUTPUT,
LLM_TENSOR_OUTPUT_NORM, LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_ROPE_FREQS, LLM_TENSOR_ROPE_FREQS,
LLM_TENSOR_ROPE_FACTORS_LONG,
LLM_TENSOR_ROPE_FACTORS_SHORT,
LLM_TENSOR_ATTN_Q, LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_K, LLM_TENSOR_ATTN_K,
LLM_TENSOR_ATTN_V, LLM_TENSOR_ATTN_V,
@ -825,18 +829,20 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
{ {
LLM_ARCH_PHI3, LLM_ARCH_PHI3,
{ {
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" }, { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" }, { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" }, { LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, { LLM_TENSOR_ROPE_FACTORS_LONG, "rope_factors_long" },
{ LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, { LLM_TENSOR_ROPE_FACTORS_SHORT, "rope_factors_short" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
}, },
}, },
{ {
@ -1776,6 +1782,7 @@ struct llama_hparams {
float f_norm_eps; float f_norm_eps;
float f_norm_rms_eps; float f_norm_rms_eps;
float rope_attn_factor = 1.0f;
float rope_freq_base_train; float rope_freq_base_train;
float rope_freq_scale_train; float rope_freq_scale_train;
uint32_t n_yarn_orig_ctx; uint32_t n_yarn_orig_ctx;
@ -1824,6 +1831,7 @@ struct llama_hparams {
if (!is_float_close(this->f_norm_eps, other.f_norm_eps, EPSILON)) return true; if (!is_float_close(this->f_norm_eps, other.f_norm_eps, EPSILON)) return true;
if (!is_float_close(this->f_norm_rms_eps, other.f_norm_rms_eps, EPSILON)) return true; if (!is_float_close(this->f_norm_rms_eps, other.f_norm_rms_eps, EPSILON)) return true;
if (!is_float_close(this->rope_attn_factor, other.rope_attn_factor, EPSILON)) return true;
if (!is_float_close(this->rope_freq_base_train, other.rope_freq_base_train, EPSILON)) return true; if (!is_float_close(this->rope_freq_base_train, other.rope_freq_base_train, EPSILON)) return true;
if (!is_float_close(this->rope_freq_scale_train, other.rope_freq_scale_train, EPSILON)) return true; if (!is_float_close(this->rope_freq_scale_train, other.rope_freq_scale_train, EPSILON)) return true;
@ -2133,6 +2141,10 @@ struct llama_model {
struct ggml_tensor * output; struct ggml_tensor * output;
struct ggml_tensor * output_b; struct ggml_tensor * output_b;
// long rope factors
struct ggml_tensor * rope_long = nullptr;
struct ggml_tensor * rope_short = nullptr;
std::vector<llama_layer> layers; std::vector<llama_layer> layers;
llama_split_mode split_mode; llama_split_mode split_mode;
@ -3337,6 +3349,39 @@ struct llama_model_loader {
return get_arr_n(llm_kv(kid), result, required); return get_arr_n(llm_kv(kid), result, required);
} }
template<typename T>
bool get_arr(const std::string & key, std::vector<T> & result, const bool required = true) {
const int kid = gguf_find_key(meta, key.c_str());
if (kid < 0) {
if (required) {
throw std::runtime_error(format("key not found in model: %s", key.c_str()));
}
return false;
}
struct GGUFMeta::ArrayInfo arr_info =
GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta, kid);
if (arr_info.gt != GGUF_TYPE_FLOAT32 && arr_info.gt != GGUF_TYPE_INT32) {
throw std::runtime_error(format("%s is not a float32 or int32 array", key.c_str()));
}
// GGML_ASSERT(gguf_type_size(arr_info.gt) == sizeof(T));
GGML_ASSERT((arr_info.gt != GGUF_TYPE_FLOAT32 || std::is_same<T, float>::value));
GGML_ASSERT((arr_info.gt != GGUF_TYPE_INT32 || std::is_same<T, int>::value));
result.resize(arr_info.length);
result.assign((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length);
return true;
}
template<typename T>
bool get_arr(const enum llm_kv kid, T& result, const bool required = true) {
return get_arr(llm_kv(kid), result, required);
}
template<typename T> template<typename T>
bool get_key(const std::string & key, T & result, const bool required = true) { bool get_key(const std::string & key, T & result, const bool required = true) {
auto it = kv_overrides.find(key); auto it = kv_overrides.find(key);
@ -3893,6 +3938,8 @@ static void llm_load_hparams(
} }
hparams.rope_freq_scale_train = ropescale == 0.0f ? 1.0f : 1.0f/ropescale; hparams.rope_freq_scale_train = ropescale == 0.0f ? 1.0f : 1.0f/ropescale;
ml.get_key(LLM_KV_ROPE_SCALING_ATTN_FACTOR, hparams.rope_attn_factor, false);
// sanity check for n_rot (optional) // sanity check for n_rot (optional)
{ {
hparams.n_rot = (hparams.n_head == 0) ? 0 : hparams.n_embd / hparams.n_head; hparams.n_rot = (hparams.n_head == 0) ? 0 : hparams.n_embd / hparams.n_head;
@ -4950,6 +4997,7 @@ static bool llm_load_tensors(
// create tensors for the weights // create tensors for the weights
{ {
const int64_t n_embd = hparams.n_embd; const int64_t n_embd = hparams.n_embd;
const int64_t n_embd_head = n_embd / hparams.n_head;
const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa();
const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa();
const int64_t n_embd_gqa = n_embd_v_gqa; const int64_t n_embd_gqa = n_embd_v_gqa;
@ -5661,6 +5709,9 @@ static bool llm_load_tensors(
{ {
model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }); model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab });
model.rope_long = ml.create_tensor(ctx_input, tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight"), { n_embd_head/2 }, false);
model.rope_short = ml.create_tensor(ctx_input, tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight"), { n_embd_head/2 }, false);
// output // output
{ {
model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }); model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd });
@ -5671,12 +5722,12 @@ static bool llm_load_tensors(
ggml_context* ctx_layer = ctx_for_layer(i); ggml_context* ctx_layer = ctx_for_layer(i);
ggml_context* ctx_split = ctx_for_layer_split(i); ggml_context* ctx_split = ctx_for_layer_split(i);
auto& layer = model.layers[i]; auto & layer = model.layers[i];
layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }); layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd });
layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, false); layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, false);
layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd, n_embd }); layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd, n_embd });
layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }); layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd });
@ -6891,17 +6942,20 @@ struct llm_build_context {
cb(lctx.inp_K_shift, "K_shift", -1); cb(lctx.inp_K_shift, "K_shift", -1);
ggml_set_input(lctx.inp_K_shift); ggml_set_input(lctx.inp_K_shift);
struct ggml_tensor * rope_factors = build_rope_factors();
for (int il = 0; il < n_layer; ++il) { for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * tmp = struct ggml_tensor * tmp =
// we rotate only the first n_rot dimensions // we rotate only the first n_rot dimensions
ggml_rope_custom_inplace(ctx0, ggml_rope_ext_inplace(ctx0,
ggml_view_3d(ctx0, kv_self.k_l[il], ggml_view_3d(ctx0, kv_self.k_l[il],
n_embd_head_k, n_head_kv, n_ctx, n_embd_head_k, n_head_kv, n_ctx,
ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k), ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k),
ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa), ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa),
0), 0),
lctx.inp_K_shift, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, lctx.inp_K_shift, rope_factors, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow); ext_factor, attn_factor, beta_fast, beta_slow);
cb(tmp, "K_shifted", il); cb(tmp, "K_shifted", il);
ggml_build_forward_expand(gf, tmp); ggml_build_forward_expand(gf, tmp);
} }
@ -7004,6 +7058,17 @@ struct llm_build_context {
return lctx.inp_pos; return lctx.inp_pos;
} }
struct ggml_tensor * build_rope_factors() {
// choose long/short freq factors based on the context size
const auto n_ctx_pre_seq = cparams.n_ctx / cparams.n_seq_max;
if (n_ctx_pre_seq > hparams.n_yarn_orig_ctx) {
return model.rope_long;
}
return model.rope_short;
}
struct ggml_tensor * build_inp_out_ids() { struct ggml_tensor * build_inp_out_ids() {
lctx.inp_out_ids = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_outputs); lctx.inp_out_ids = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_outputs);
cb(lctx.inp_out_ids, "inp_out_ids", -1); cb(lctx.inp_out_ids, "inp_out_ids", -1);
@ -7111,15 +7176,15 @@ struct llm_build_context {
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
} }
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
@ -7241,13 +7306,13 @@ struct llm_build_context {
switch (model.type) { switch (model.type) {
case MODEL_7B: case MODEL_7B:
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
@ -7353,15 +7418,15 @@ struct llm_build_context {
struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
@ -7474,14 +7539,14 @@ struct llm_build_context {
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
// using mode = 2 for neox mode // using mode = 2 for neox mode
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Kcur, "Kcur", il); cb(Kcur, "Kcur", il);
@ -7597,15 +7662,15 @@ struct llm_build_context {
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
} }
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
@ -7749,15 +7814,15 @@ struct llm_build_context {
cb(Kcur, "Kcur", il); cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
@ -8102,15 +8167,15 @@ struct llm_build_context {
cb(Kcur, "Kcur", il); cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
@ -8542,15 +8607,15 @@ struct llm_build_context {
} }
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, ctx0, Qcur, inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, ctx0, Kcur, inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
@ -8662,14 +8727,14 @@ struct llm_build_context {
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
// using mode = 2 for neox mode // using mode = 2 for neox mode
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Kcur, "Kcur", il); cb(Kcur, "Kcur", il);
@ -8773,15 +8838,15 @@ struct llm_build_context {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
@ -8887,15 +8952,15 @@ struct llm_build_context {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
@ -9039,8 +9104,8 @@ struct llm_build_context {
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
@ -9050,8 +9115,8 @@ struct llm_build_context {
Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head))); Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head)));
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Kcur, "Kcur", il); cb(Kcur, "Kcur", il);
@ -9122,6 +9187,9 @@ struct llm_build_context {
// KQ_mask (mask for 1 head, it will be broadcasted to all heads) // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
// rope freq factors for 128k context
struct ggml_tensor * rope_factors = build_rope_factors();
for (int il = 0; il < n_layer; ++il) { for (int il = 0; il < n_layer; ++il) {
auto residual = inpL; auto residual = inpL;
@ -9158,8 +9226,8 @@ struct llm_build_context {
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
@ -9167,8 +9235,8 @@ struct llm_build_context {
Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head))); Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head)));
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Kcur, "Kcur", il); cb(Kcur, "Kcur", il);
@ -9274,14 +9342,14 @@ struct llm_build_context {
struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_rot, n_head, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Qcur, n_rot, n_head, n_tokens), inp_pos, nullptr,
n_embd_head, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_embd_head, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow); ext_factor, attn_factor, beta_fast, beta_slow);
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens), inp_pos, nullptr,
n_embd_head, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_embd_head, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow); ext_factor, attn_factor, beta_fast, beta_slow);
cb(Kcur, "Kcur", il); cb(Kcur, "Kcur", il);
@ -9482,15 +9550,15 @@ struct llm_build_context {
cb(tmpk, "tmpk", il); cb(tmpk, "tmpk", il);
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
struct ggml_tensor * Qcur = ggml_rope_custom( struct ggml_tensor * Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
struct ggml_tensor * Kcur = ggml_rope_custom( struct ggml_tensor * Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
@ -9598,15 +9666,15 @@ struct llm_build_context {
// cb(Vcur, "Vcur", il); // cb(Vcur, "Vcur", il);
// } // }
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
@ -9715,15 +9783,15 @@ struct llm_build_context {
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
} }
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
@ -9845,15 +9913,15 @@ struct llm_build_context {
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
} }
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
@ -9965,8 +10033,8 @@ struct llm_build_context {
struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head, n_tokens), inp_pos, nullptr,
n_embd_head_k, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_embd_head_k, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow); ext_factor, attn_factor, beta_fast, beta_slow);
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
@ -9974,8 +10042,8 @@ struct llm_build_context {
Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k))); Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));
cb(Qcur, "Qcur_scaled", il); cb(Qcur, "Qcur_scaled", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr,
n_embd_head_k, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_embd_head_k, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow); ext_factor, attn_factor, beta_fast, beta_slow);
cb(Kcur, "Kcur", il); cb(Kcur, "Kcur", il);
@ -10085,15 +10153,15 @@ struct llm_build_context {
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
} }
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
@ -10375,15 +10443,15 @@ struct llm_build_context {
cb(Kcur, "Kcur", il); cb(Kcur, "Kcur", il);
} }
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
@ -10506,15 +10574,15 @@ struct llm_build_context {
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
} }
Qcur = ggml_rope_custom( Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom( Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
@ -12789,15 +12857,16 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
// tokenizer.encode('', add_special_tokens=True) returns [1] // tokenizer.encode('', add_special_tokens=True) returns [1]
// tokenizer.encode('', add_special_tokens=False) returns [] // tokenizer.encode('', add_special_tokens=False) returns []
if (add_special && vocab.special_add_bos != 0) {
GGML_ASSERT(vocab.special_bos_id != -1);
output.push_back(vocab.special_bos_id);
}
static const bool rtrim = true; //TODO: as param static const bool rtrim = true; //TODO: as param
bool is_prev_special = false; bool is_prev_special = false;
bool special_token_rtrim = false; bool special_token_rtrim = false;
if (add_special && vocab.special_add_bos != 0) {
GGML_ASSERT(vocab.special_bos_id != -1);
output.push_back(vocab.special_bos_id);
is_prev_special = true;
}
for (const auto & fragment : fragment_buffer) { for (const auto & fragment : fragment_buffer) {
if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) { if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
// without adding this leading whitespace, we do not get the same results as the original tokenizer // without adding this leading whitespace, we do not get the same results as the original tokenizer
@ -15719,6 +15788,7 @@ struct llama_context * llama_new_context_with_model(
cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_YARN ? 1.0f : 0.0f; cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_YARN ? 1.0f : 0.0f;
} }
cparams.yarn_attn_factor *= hparams.rope_attn_factor;
cparams.causal_attn = hparams.causal_attn; cparams.causal_attn = hparams.causal_attn;
if (cparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) { if (cparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) {

34
tests/test-tokenizer-0.sh
View file

@ -1,34 +0,0 @@
#!/bin/bash
#
# Usage:
#
# test-tokenizer-0.sh <name> <input>
#
if [ $# -ne 2 ]; then
printf "Usage: $0 <name> <input>\n"
exit 1
fi
name=$1
input=$2
make -j tests/test-tokenizer-0
printf "Testing %s on %s ...\n" $name $input
python3 ./tests/test-tokenizer-0.py ./models/tokenizers/$name --fname-tok $input > /tmp/test-tokenizer-0-$name-py.log 2>&1
cat /tmp/test-tokenizer-0-$name-py.log | grep "tokenized in"
./tests/test-tokenizer-0 ./models/ggml-vocab-$name.gguf $input > /tmp/test-tokenizer-0-$name-cpp.log 2>&1
cat /tmp/test-tokenizer-0-$name-cpp.log | grep "tokenized in"
diff $input.tok $input.tokcpp > /dev/null 2>&1
if [ $? -eq 0 ]; then
printf "Tokenization is correct!\n"
else
diff $input.tok $input.tokcpp | head -n 32
printf "Tokenization differs!\n"
fi

331
tests/test-tokenizer-random.py
View file

@ -1,331 +0,0 @@
# Test libllama tokenizer == AutoTokenizer.
# Brute force random words/text generation.
#
# Sample usage:
#
# python3 tests/test-tokenizer-random.py ./models/ggml-vocab-llama-bpe.gguf ./models/tokenizers/llama-bpe
#
import time
import logging
import argparse
import subprocess
import random
from typing import Callable, Iterator
import cffi
from transformers import AutoTokenizer
logger = logging.getLogger("test-tokenizer-random-bpe")
class LibLlama:
DEFAULT_PATH_LLAMA_H = "./llama.h"
DEFAULT_PATH_LIBLLAMA = "./build/libllama.so" # CMakeLists.txt: BUILD_SHARED_LIBS ON
def __init__(self, path_llama_h: str = None, path_libllama: str = None):
path_llama_h = path_llama_h or self.DEFAULT_PATH_LLAMA_H
path_libllama = path_libllama or self.DEFAULT_PATH_LIBLLAMA
(self.ffi, self.lib) = self._load_libllama_cffi(path_llama_h, path_libllama)
self.lib.llama_backend_init()
def _load_libllama_cffi(self, path_llama_h: str, path_libllama: str):
cmd = ["gcc", "-E", "-P", "-D__restrict=", "-D__attribute__(x)=", "-D__asm__(x)=", path_llama_h]
res = subprocess.run(cmd, stdout=subprocess.PIPE)
assert (res.returncode == 0)
source = res.stdout.decode()
ffi = cffi.FFI()
if True: # workarounds for pycparser
source = "typedef struct { } __builtin_va_list;" + "\n" + source
source = source.replace("sizeof (int)", str(ffi.sizeof("int")))
source = source.replace("sizeof (void *)", str(ffi.sizeof("void*")))
source = source.replace("sizeof (size_t)", str(ffi.sizeof("size_t")))
source = source.replace("sizeof(int32_t)", str(ffi.sizeof("int32_t")))
ffi.cdef(source, override=True)
lib = ffi.dlopen(path_libllama)
return (ffi, lib)
def model_default_params(self, **kwargs):
mparams = self.lib.llama_model_default_params()
for k, v in kwargs.items():
setattr(mparams, k, v)
return mparams
def context_default_params(self, **kwargs):
cparams = self.lib.llama_context_default_params()
for k, v in kwargs.items():
setattr(cparams, k, v)
return cparams
class LibLlamaModel:
def __init__(self, libllama: LibLlama, path_model: str, mparams={}, cparams={}):
self.lib = libllama.lib
self.ffi = libllama.ffi
if isinstance(mparams, dict):
mparams = libllama.model_default_params(**mparams)
self.model = self.lib.llama_load_model_from_file(path_model.encode(), mparams)
if not self.model:
raise RuntimeError("error: failed to load model '%s'" % path_model)
if isinstance(cparams, dict):
cparams = libllama.context_default_params(**cparams)
self.ctx = self.lib.llama_new_context_with_model(self.model, cparams)
if not self.ctx:
raise RuntimeError("error: failed to create context for model '%s'" % path_model)
n_tokens_max = self.lib.llama_n_ctx(self.ctx)
self.token_ids = self.ffi.new("llama_token[]", n_tokens_max)
def free(self):
if self.ctx:
self.lib.llama_free(self.ctx)
if self.model:
self.lib.llama_free_model(self.model)
self.ctx = None
self.model = None
self.lib = None
def tokenize(self, text: str, n_tokens_max: int = 0, add_special: bool = False, parse_special: bool = False) -> list[int]:
n_tokens_max = n_tokens_max if n_tokens_max > 0 else len(self.token_ids)
text = text.encode("utf-8")
num = self.lib.llama_tokenize(self.model, text, len(text), self.token_ids, n_tokens_max, add_special, parse_special)
if num < 0:
return []
return list(self.token_ids[0:num])
def generator_custom_text() -> Iterator[str]:
"""General tests"""
yield from [
"",
" ",
" ",
" ",
"\t",
"\n",
"\n\n",
"\n\n\n",
"\t\n",
"Hello world",
" Hello world",
"Hello World",
" Hello World",
" Hello World!",
"Hello, world!",
" Hello, world!",
" this is 🦙.cpp",
"w048 7tuijk dsdfhu",
"нещо на Български",
"កាន់តែពិសេសអាចខលចេញ",
"🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)",
"Hello",
" Hello",
" Hello",
" Hello",
" Hello",
" Hello\n Hello",
" (",
"\n =",
"' era",
"Hello, y'all! How are you 😁 ?我想在apple工作1314151天～",
"3",
"33",
"333",
"3333",
"33333",
"333333",
"3333333",
"33333333",
"333333333",
]
def generator_custom_text_edge_cases() -> Iterator[str]:
"""Edge cases found while debugging"""
yield from [
'\x1f-a', # unicode_ranges_control, {0x00001C, 0x00001F}
'¼-a', # unicode_ranges_digit, 0x00BC
'½-a', # unicode_ranges_digit, 0x00BD
'¾-a', # unicode_ranges_digit, 0x00BE
'a 〇b', # unicode_ranges_digit, 0x3007
'Ⅵ-a', # unicode_ranges_digit, {0x00002150, 0x0000218F} // Number Forms
'\uFEFF//', # unicode_ranges_control, 0xFEFF (BOM)
'Cửa Việt', # llama-3, ignore_merges = true
'<s>a', # Phi-3 fail
'<unk><|endoftext|><s>' # Phi-3 fail
'a\na', # TODO: Bert fail
]
def generator_random_special_tokens(special_tokens:list[str], iterations=100) -> Iterator[str]:
special_tokens = set(special_tokens)
special_tokens.update([" ", "\n", "\t", "-", "!", "one", "1", "<s>", "</s>"])
special_tokens = list(sorted(special_tokens))
rand = random.Random()
for m in range(iterations):
rand.seed(m)
words = rand.choices(special_tokens, k=500)
yield "".join(words)
def generator_vocab_words(vocab: list[str]) -> Iterator[str]:
"""Brute force check all vocab words"""
yield from vocab
def generator_random_chars(iterations=100) -> Iterator[str]:
"""Brute force random text with simple characters"""
WHITESPACES = list(" " * 20 + "\n" * 5 + "\r\n" * 5 + "\t" * 5)
CHARS = list(sorted(set("""
ABCDEFGHIJKLMNOPQRSTUVWXYZ
abcdefghijklmnopqrstuvwxyz
ÁÉÍÓÚÀÈÌÒÙÂÊÎÔÛÄËÏÖÜ
áéíóúàèìòùâêîôûäëïöü
.-,*/-+ª!"·$%&/()=?¿[]{}<>\\|@#~½¬~;:_
""")))
rand = random.Random()
for m in range(iterations):
rand.seed(m)
text = []
num_words = rand.randint(300, 400)
for i in range(num_words):
k = rand.randint(1, 7)
word = rand.choices(CHARS, k=k)
space = rand.choice(WHITESPACES)
text.append("".join(word) + space)
yield "".join(text)
def generator_random_vocab_chars(vocab: list[str], iterations=100) -> Iterator[str]:
"""Brute force random text with vocab characters"""
vocab_chars = set()
for word in vocab:
vocab_chars.update(word)
vocab_chars = list(sorted(vocab_chars))
rand = random.Random()
for m in range(iterations):
rand.seed(m)
text = rand.choices(vocab_chars, k=1024)
yield "".join(text)
def generator_random_vocab_words(vocab: list[str], iterations=100) -> Iterator[str]:
"""Brute force random text from vocab words"""
vocab = [w.strip() for w in vocab]
yield from vocab
rand = random.Random()
for m in range(iterations):
rand.seed(m)
text = []
num_words = rand.randint(300, 400)
for i in range(num_words):
k = rand.randint(1, 3)
words = rand.choices(vocab, k=k)
sep = rand.choice(" \n\r\t")
text.append("".join(words) + sep)
yield "".join(text)
def generator_random_bytes(iterations=100) -> Iterator[str]:
"""Brute force random bytes"""
WHITESPACES = list(" " * 20 + "\n" * 5 + "\r\n" * 5 + "\t" * 5)
rand = random.Random()
for m in range(iterations):
rand.seed(m)
text = []
num_words = rand.randint(300, 400)
for i in range(num_words):
k = rand.randint(1, 8)
word = [chr(r) for r in rand.randbytes(k) if r]
word.append(rand.choice(WHITESPACES))
text.append("".join(word))
yield "".join(text)
def test_compare_tokenizer(func_tokenize1: Callable, func_tokenize2: Callable, generator: Iterator[str]):
def find_first_mismatch(ids1: list[int], ids2: list[int]):
for i, (a, b) in enumerate(zip(ids1, ids2)):
if a != b:
return i
if len(ids1) == len(ids2):
return -1
return min(len(ids1), len(ids2))
t0 = time.perf_counter()
logger.info("%s: %s" % (generator.__name__, "ini"))
for text in generator:
ids1 = func_tokenize1(text)
ids2 = func_tokenize2(text)
if ids1 != ids2:
i = find_first_mismatch(ids1, ids2)
ids1 = list(ids1)[max(0, i - 2) : i + 2 + 1]
ids2 = list(ids2)[max(0, i - 2) : i + 2 + 1]
logger.info(" TokenIDs: " + str(ids1))
logger.info(" Expected: " + str(ids2))
raise Exception()
t1 = time.perf_counter()
logger.info("%s: end, time: %.3f secs" % (generator.__name__, t1 - t0))
def main(argv: list[str] = None):
parser = argparse.ArgumentParser()
parser.add_argument("vocab_file", help="path to vocab 'gguf' file")
parser.add_argument("dir_tokenizer", help="directory containing 'tokenizer.model' file")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
args = parser.parse_args(argv)
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
model = LibLlamaModel(LibLlama(), args.vocab_file, mparams=dict(vocab_only=True), cparams=dict(n_ctx=4096))
tokenizer = AutoTokenizer.from_pretrained(args.dir_tokenizer)
def func_tokenize2(text: str):
return tokenizer.encode(text, add_special_tokens=False)
parse_special = all(len(func_tokenize2(t)) == 1 for t in tokenizer.all_special_tokens)
def func_tokenize1(text: str):
return model.tokenize(text, add_special=False, parse_special=parse_special)
vocab = list(sorted(tokenizer.batch_decode(list(tokenizer.get_vocab().values()), skip_special_tokens=True)))
test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text())
test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text_edge_cases())
test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_special_tokens(tokenizer.all_special_tokens, 10_000))
test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_vocab_words(vocab))
test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_chars(10_000))
test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_vocab_chars(vocab, 10_000))
test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_vocab_words(vocab, 5_000))
# test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_bytes(10_000)) # FAIL
model.free()
if __name__ == "__main__":
# main()
path_tokenizers = "./models/tokenizers/"
path_vocab_format = "./models/ggml-vocab-%s.gguf"
# import os
# tokenizers = os.listdir(path_tokenizers)
tokenizers = [
"llama-spm", # SPM
"phi-3", # SPM
]
for tokenizer in tokenizers:
print("\n" + "=" * 50 + "\n" + tokenizer + "\n") # noqa
vocab_file = path_vocab_format % tokenizer
dir_tokenizer = path_tokenizers + "/" + tokenizer
main([vocab_file, dir_tokenizer, "--verbose"])