vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into xsn/chat_template_prefix_postfix

Browse source
This commit is contained in:
ngxson 2024-04-24 18:12:13 +02:00
commit ada54292c6
29 changed files with 1188 additions and 936 deletions
Download patch file Download diff file Expand all files Collapse all files

2
.github/workflows/bench.yml vendored
View file

@ -32,7 +32,7 @@ on:
- cron: '04 2 * * *'
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}-${{ github.event.inputs.sha }}
group: ${{ github.workflow }}-${{ github.ref || github.run_id }}-${{ github.event.inputs.sha }}
cancel-in-progress: true
jobs:

2
.github/workflows/server.yml vendored
View file

@ -23,7 +23,7 @@ on:
- cron: '2 4 * * *'
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
group: ${{ github.workflow }}-${{ github.ref || github.run_id }}
cancel-in-progress: true
jobs:

16
CMakeLists.txt
View file

@ -43,17 +43,11 @@ else()
set(LLAMA_METAL_DEFAULT OFF)
endif()
# TODO: fix this for Android CI
# https://github.com/ggerganov/llama.cpp/pull/6716#issuecomment-2061509191
#if (CMAKE_SYSTEM_NAME MATCHES "ANDROID")
# set(LLAMA_LLAMAFILE_DEFAULT OFF)
#else()
# set(LLAMA_LLAMAFILE_DEFAULT ON)
#endif()
# TODO: temporary disable until MoE is fixed
# https://github.com/ggerganov/llama.cpp/pull/6716
set(LLAMA_LLAMAFILE_DEFAULT OFF)
if (CMAKE_SYSTEM_NAME MATCHES "ANDROID")
set(LLAMA_LLAMAFILE_DEFAULT OFF)
else()
set(LLAMA_LLAMAFILE_DEFAULT ON)
endif()
# general
option(BUILD_SHARED_LIBS "build shared libraries" OFF)

4
Makefile
View file

@ -384,10 +384,6 @@ ifdef LLAMA_OPENBLAS
MK_LDFLAGS += $(shell pkg-config --libs openblas)
endif # LLAMA_OPENBLAS
# TODO: temporary disable until MoE is fixed
# https://github.com/ggerganov/llama.cpp/pull/6716
LLAMA_NO_LLAMAFILE := 1
ifndef LLAMA_NO_LLAMAFILE
MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
OBJS += sgemm.o

20
README-sycl.md
View file

@ -229,12 +229,12 @@ source /opt/intel/oneapi/setvars.sh
# Build LLAMA with MKL BLAS acceleration for intel GPU
mkdir -p build && cd build
# Option 1: Use FP16 for better performance in long-prompt inference
#cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
# Option 2: Use FP32 by default
# Option 1: Use FP32 (recommended for better performance in most cases)
cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
# Option 2: Use FP16
cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
#build all binary
cmake --build . --config Release -j -v
```
@ -250,12 +250,12 @@ export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
# Build LLAMA with Nvidia BLAS acceleration through SYCL
mkdir -p build && cd build
# Option 1: Use FP16 for better performance in long-prompt inference
cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
# Option 2: Use FP32 by default
# Option 1: Use FP32 (recommended for better performance in most cases)
cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
# Option 2: Use FP16
cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
#build all binary
cmake --build . --config Release -j -v
@ -416,6 +416,10 @@ mkdir -p build
cd build
@call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
# Option 1: Use FP32 (recommended for better performance in most cases)
cmake -G "MinGW Makefiles" .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Release
# Option 2: Or FP16
cmake -G "MinGW Makefiles" .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
make -j

8
common/common.cpp
View file

@ -242,7 +242,9 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
invalid_param = true;
return true;
}
// This is temporary, in the future the samplign state will be moved fully to llama_sampling_context.
params.seed = std::stoul(argv[i]);
sparams.seed = std::stoul(argv[i]);
return true;
}
if (arg == "-t" || arg == "--threads") {
@ -2326,12 +2328,12 @@ std::vector<llama_token> llama_tokenize(
return result;
}
std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
std::vector<char> result(8, 0);
const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), true);
const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special);
if (n_tokens < 0) {
result.resize(-n_tokens);
int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), true);
int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special);
GGML_ASSERT(check == -n_tokens);
} else {
result.resize(n_tokens);

9
common/common.h
View file

@ -86,8 +86,8 @@ struct gpt_params {
ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
// // sampling parameters
struct llama_sampling_params sparams;
@ -237,11 +237,12 @@ std::vector<llama_token> llama_tokenize(
bool add_special,
bool parse_special = false);
// tokenizes a token into a piece
// tokenizes a token into a piece, optionally renders special/control tokens
// should work similar to Python's `tokenizer.id_to_piece`
std::string llama_token_to_piece(
const struct llama_context * ctx,
llama_token token);
llama_token token,
bool special = true);
// TODO: these should be moved in llama.h C-style API under single `llama_detokenize` function
// that takes into account the tokenizer type and decides how to handle the leading space

13
common/sampling.cpp
View file

@ -1,4 +1,6 @@
#define LLAMA_API_INTERNAL
#include "sampling.h"
#include <random>
struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params) {
struct llama_sampling_context * result = new llama_sampling_context();
@ -33,6 +35,8 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
result->prev.resize(params.n_prev);
llama_sampling_set_rng_seed(result, params.seed);
return result;
}
@ -62,6 +66,13 @@ void llama_sampling_reset(llama_sampling_context * ctx) {
ctx->cur.clear();
}
void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed) {
if (seed == LLAMA_DEFAULT_SEED) {
seed = time(NULL);
}
ctx->rng.seed(seed);
}
void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst) {
if (dst->grammar) {
llama_grammar_free(dst->grammar);
@ -203,7 +214,7 @@ static llama_token llama_sampling_sample_impl(
sampler_queue(ctx_main, params, cur_p, min_keep);
id = llama_sample_token(ctx_main, &cur_p);
id = llama_sample_token_with_rng(ctx_main, &cur_p, ctx_sampling->rng);
//{
// const int n_top = 10;

47
common/sampling.h
View file

@ -4,9 +4,10 @@
#include "grammar-parser.h"
#include <random>
#include <string>
#include <vector>
#include <unordered_map>
#include <vector>
// sampler types
enum class llama_sampler_type : char {
@ -20,25 +21,26 @@ enum class llama_sampler_type : char {
// sampling parameters
typedef struct llama_sampling_params {
int32_t n_prev = 64; // number of previous tokens to remember
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
int32_t min_keep = 0; // 0 = disabled, otherwise samplers should return at least min_keep tokens
int32_t top_k = 40; // <= 0 to use vocab size
float top_p = 0.95f; // 1.0 = disabled
float min_p = 0.05f; // 0.0 = disabled
float tfs_z = 1.00f; // 1.0 = disabled
float typical_p = 1.00f; // 1.0 = disabled
float temp = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
float dynatemp_range = 0.00f; // 0.0 = disabled
float dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
int32_t penalty_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size)
float penalty_repeat = 1.00f; // 1.0 = disabled
float penalty_freq = 0.00f; // 0.0 = disabled
float penalty_present = 0.00f; // 0.0 = disabled
int32_t mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
float mirostat_tau = 5.00f; // target entropy
float mirostat_eta = 0.10f; // learning rate
bool penalize_nl = false; // consider newlines as a repeatable token
int32_t n_prev = 64; // number of previous tokens to remember
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
int32_t min_keep = 0; // 0 = disabled, otherwise samplers should return at least min_keep tokens
int32_t top_k = 40; // <= 0 to use vocab size
float top_p = 0.95f; // 1.0 = disabled
float min_p = 0.05f; // 0.0 = disabled
float tfs_z = 1.00f; // 1.0 = disabled
float typical_p = 1.00f; // 1.0 = disabled
float temp = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
float dynatemp_range = 0.00f; // 0.0 = disabled
float dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
int32_t penalty_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size)
float penalty_repeat = 1.00f; // 1.0 = disabled
float penalty_freq = 0.00f; // 0.0 = disabled
float penalty_present = 0.00f; // 0.0 = disabled
int32_t mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
float mirostat_tau = 5.00f; // target entropy
float mirostat_eta = 0.10f; // learning rate
bool penalize_nl = false; // consider newlines as a repeatable token
uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampling_context
std::vector<llama_sampler_type> samplers_sequence = {
llama_sampler_type::TOP_K,
@ -79,6 +81,8 @@ struct llama_sampling_context {
// TODO: replace with ring-buffer
std::vector<llama_token> prev;
std::vector<llama_token_data> cur;
std::mt19937 rng;
};
#include "common.h"
@ -93,6 +97,9 @@ void llama_sampling_free(struct llama_sampling_context * ctx);
// - reset grammar
void llama_sampling_reset(llama_sampling_context * ctx);
// Set the sampler seed
void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed);
// Copy the sampler context
void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst);

101
convert-hf-to-gguf.py
View file

@ -363,6 +363,16 @@ class Model(ABC):
scores.append(-1000.0)
toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
if vocab_size > len(tokens):
pad_count = vocab_size - len(tokens)
print(
f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]"
)
for i in range(1, pad_count + 1):
tokens.append(f"[PAD{i}]")
scores.append(-1000.0)
toktypes.append(SentencePieceTokenTypes.UNUSED)
assert len(tokens) == vocab_size
self.gguf_writer.add_tokenizer_model("llama")
@ -1789,6 +1799,12 @@ class QwenModel(Model):
class Qwen2Model(Model):
model_arch = gguf.MODEL_ARCH.QWEN2
def set_vocab(self):
try:
self._set_vocab_sentencepiece()
except FileNotFoundError:
self._set_vocab_gpt2()
@Model.register("Qwen2MoeForCausalLM")
class Qwen2MoeModel(Model):
@ -1979,6 +1995,91 @@ class Phi2Model(Model):
self.gguf_writer.add_add_bos_token(False)
@Model.register("Phi3ForCausalLM")
class Phi3MiniModel(Model):
model_arch = gguf.MODEL_ARCH.PHI3
def set_vocab(self):
from sentencepiece import SentencePieceProcessor
tokenizer_path = self.dir_model / 'tokenizer.model'
if not tokenizer_path.is_file():
print(f'Error: Missing {tokenizer_path}', file=sys.stderr)
sys.exit(1)
tokenizer = SentencePieceProcessor(str(tokenizer_path))
vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
scores: list[float] = [-10000.0] * vocab_size
toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
for token_id in range(tokenizer.vocab_size()):
piece = tokenizer.id_to_piece(token_id)
text = piece.encode("utf-8")
score = tokenizer.get_score(token_id)
toktype = SentencePieceTokenTypes.NORMAL
if tokenizer.is_unknown(token_id):
toktype = SentencePieceTokenTypes.UNKNOWN
elif tokenizer.is_control(token_id):
toktype = SentencePieceTokenTypes.CONTROL
elif tokenizer.is_unused(token_id):
toktype = SentencePieceTokenTypes.UNUSED
elif tokenizer.is_byte(token_id):
toktype = SentencePieceTokenTypes.BYTE
tokens[token_id] = text
scores[token_id] = score
toktypes[token_id] = toktype
added_tokens_file = self.dir_model / 'added_tokens.json'
if added_tokens_file.is_file():
with open(added_tokens_file, "r", encoding="utf-8") as f:
added_tokens_json = json.load(f)
for key in added_tokens_json:
token_id = added_tokens_json[key]
if (token_id >= vocab_size):
print(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
continue
tokens[token_id] = key.encode("utf-8")
scores[token_id] = -1000.0
toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
self.gguf_writer.add_tokenizer_model("llama")
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_scores(scores)
self.gguf_writer.add_token_types(toktypes)
special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
special_vocab.add_to_gguf(self.gguf_writer)
def set_gguf_parameters(self):
block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
rot_pct = 1.0
n_embd = self.find_hparam(["hidden_size", "n_embd"])
n_head = self.find_hparam(["num_attention_heads", "n_head"])
rms_eps = self.find_hparam(["rms_norm_eps"])
self.gguf_writer.add_name("Phi3")
self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"]))
self.gguf_writer.add_embedding_length(n_embd)
self.gguf_writer.add_feed_forward_length(8192)
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_head_count(n_head)
self.gguf_writer.add_head_count_kv(n_head)
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_file_type(self.ftype)
@Model.register("PlamoForCausalLM")
class PlamoModel(Model):
model_arch = gguf.MODEL_ARCH.PLAMO

1
examples/lookup/lookup-stats.cpp
View file

@ -30,7 +30,6 @@ int main(int argc, char ** argv){
// load the model
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_set_rng_seed(ctx, params.seed);
GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
// tokenize the prompt

1
examples/lookup/lookup.cpp
View file

@ -38,7 +38,6 @@ int main(int argc, char ** argv){
// load the model
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_set_rng_seed(ctx, params.seed);
GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
// tokenize the prompt

1
examples/main/main.cpp
View file

@ -240,7 +240,6 @@ int main(int argc, char ** argv) {
return 1;
}
session_tokens.resize(n_token_count_out);
llama_set_rng_seed(ctx, params.seed);
LOG_TEE("%s: loaded a session with prompt size of %d tokens\n", __func__, (int)session_tokens.size());
}
}

10
examples/server/public/index.html
View file

@ -881,11 +881,11 @@
.replace(/&/g, '&amp;')
.replace(/</g, '&lt;')
.replace(/>/g, '&gt;')
.replace(/^#{1,6} (.*)$/gim, '<h3>$1</h3>')
.replace(/\*\*(.*?)\*\*/g, '<strong>$1</strong>')
.replace(/__(.*?)__/g, '<strong>$1</strong>')
.replace(/\*(.*?)\*/g, '<em>$1</em>')
.replace(/_(.*?)_/g, '<em>$1</em>')
.replace(/(^|\n)#{1,6} ([^\n]*)(?=([^`]*`[^`]*`)*[^`]*$)/g, '$1<h3>$2</h3>')
.replace(/\*\*(.*?)\*\*(?=([^`]*`[^`]*`)*[^`]*$)/g, '<strong>$1</strong>')
.replace(/__(.*?)__(?=([^`]*`[^`]*`)*[^`]*$)/g, '<strong>$1</strong>')
.replace(/\*(.*?)\*(?=([^`]*`[^`]*`)*[^`]*$)/g, '<em>$1</em>')
.replace(/_(.*?)_(?=([^`]*`[^`]*`)*[^`]*$)/g, '<em>$1</em>')
.replace(/```.*?\n([\s\S]*?)```/g, '<pre><code>$1</code></pre>')
.replace(/`(.*?)`/g, '<code>$1</code>')
.replace(/\n/gim, '<br />');

5
examples/server/server.cpp
View file

@ -854,7 +854,7 @@ struct server_context {
slot.sparams.penalize_nl = json_value(data, "penalize_nl", default_sparams.penalize_nl);
slot.params.n_keep = json_value(data, "n_keep", slot.params.n_keep);
slot.params.n_discard = json_value(data, "n_discard", default_params.n_discard);
slot.params.seed = json_value(data, "seed", default_params.seed);
slot.sparams.seed = json_value(data, "seed", default_sparams.seed);
slot.sparams.n_probs = json_value(data, "n_probs", default_sparams.n_probs);
slot.sparams.min_keep = json_value(data, "min_keep", default_sparams.min_keep);
@ -1028,7 +1028,6 @@ struct server_context {
send_error(task, "Failed to parse grammar", ERROR_TYPE_INVALID_REQUEST);
return false;
}
llama_set_rng_seed(ctx, slot.params.seed);
}
slot.command = SLOT_COMMAND_LOAD_PROMPT;
@ -1118,7 +1117,7 @@ struct server_context {
bool process_token(completion_token_output & result, server_slot & slot) {
// remember which tokens were sampled - used for repetition penalties during sampling
const std::string token_str = llama_token_to_piece(ctx, result.tok);
const std::string token_str = llama_token_to_piece(ctx, result.tok, false);
slot.sampled = result.tok;
// search stop word and delete it

57
examples/server/tests/features/results.feature Normal file
View file

@ -0,0 +1,57 @@
@llama.cpp
@results
Feature: Results
Background: Server startup
Given a server listening on localhost:8080
And a model file tinyllamas/split/stories15M-00001-of-00003.gguf from HF repo ggml-org/models
And a model file test-model-00001-of-00003.gguf
And 128 as batch size
And 256 KV cache size
And 128 max tokens to predict
Scenario Outline: Multi users completion
Given <n_slots> slots
And continuous batching
Then the server is starting
Then the server is healthy
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given concurrent completion requests
Then the server is busy
Then the server is idle
And all slots are idle
Then all predictions are equal
Examples:
| n_slots |
| 1 |
| 2 |

34
examples/server/tests/features/steps/steps.py
View file

@ -61,6 +61,7 @@ def step_server_config(context, server_fqdn, server_port):
context.server_metrics = False
context.server_process = None
context.seed = None
context.draft = None
context.server_seed = None
context.user_api_key = None
context.response_format = None
@ -107,6 +108,11 @@ def step_n_gpu_layer(context, ngl):
context.n_gpu_layer = ngl
@step('{draft:d} as draft')
def step_draft(context, draft):
context.draft = draft
@step('{n_ctx:d} KV cache size')
def step_n_ctx(context, n_ctx):
context.n_ctx = n_ctx
@ -254,6 +260,15 @@ def step_n_tokens_predicted(context, predicted_n):
assert_n_tokens_predicted(context.completion, predicted_n)
@step('all predictions are equal')
@async_run_until_complete
async def step_predictions_equal(context):
n_completions = await gather_tasks_results(context)
assert n_completions >= 2, "need at least 2 completions"
assert_all_predictions_equal(context.tasks_result)
context.tasks_result = []
@step('the completion is truncated')
def step_assert_completion_truncated(context):
step_assert_completion_truncated(context, '')
@ -1020,6 +1035,23 @@ def assert_n_tokens_predicted(completion_response, expected_predicted_n=None, re
assert n_predicted == expected_predicted_n, (f'invalid number of tokens predicted:'
f' {n_predicted} <> {expected_predicted_n}')
def assert_all_predictions_equal(completion_responses):
content_0 = completion_responses[0]['content']
if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
print(f"content 0: {content_0}")
i = 1
for response in completion_responses[1:]:
content = response['content']
if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
print(f"content {i}: {content}")
assert content == content_0, "contents not equal"
i += 1
async def gather_tasks_results(context):
n_tasks = len(context.concurrent_tasks)
@ -1148,6 +1180,8 @@ def start_server_background(context):
server_args.extend(['--ubatch-size', context.n_ubatch])
if context.n_gpu_layer:
server_args.extend(['--n-gpu-layers', context.n_gpu_layer])
if context.draft is not None:
server_args.extend(['--draft', context.draft])
if context.server_continuous_batching:
server_args.append('--cont-batching')
if context.server_embeddings:

6
flake.lock generated
View file

@ -20,11 +20,11 @@
},
"nixpkgs": {
"locked": {
"lastModified": 1712791164,
"narHash": "sha256-3sbWO1mbpWsLepZGbWaMovSO7ndZeFqDSdX0hZ9nVyw=",
"lastModified": 1713537308,
"narHash": "sha256-XtTSSIB2DA6tOv+l0FhvfDMiyCmhoRbNB+0SeInZkbk=",
"owner": "NixOS",
"repo": "nixpkgs",
"rev": "1042fd8b148a9105f3c0aca3a6177fd1d9360ba5",
"rev": "5c24cf2f0a12ad855f444c30b2421d044120c66f",
"type": "github"
},
"original": {

16
ggml-alloc.c
View file

@ -371,16 +371,16 @@ struct ggml_gallocr {
};
ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs) {
ggml_gallocr_t galloc = (ggml_gallocr_t)calloc(sizeof(struct ggml_gallocr), 1);
ggml_gallocr_t galloc = (ggml_gallocr_t)calloc(1, sizeof(struct ggml_gallocr));
GGML_ASSERT(galloc != NULL);
galloc->bufts = calloc(sizeof(ggml_backend_buffer_type_t) * n_bufs, 1);
galloc->bufts = calloc(n_bufs, sizeof(ggml_backend_buffer_type_t));
GGML_ASSERT(galloc->bufts != NULL);
galloc->buffers = calloc(sizeof(ggml_backend_buffer_t) * n_bufs, 1);
galloc->buffers = calloc(n_bufs, sizeof(ggml_backend_buffer_t) * n_bufs);
GGML_ASSERT(galloc->buffers != NULL);
galloc->buf_tallocs = calloc(sizeof(struct ggml_dyn_tallocr *) * n_bufs, 1);
galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));
GGML_ASSERT(galloc->buf_tallocs != NULL);
for (int i = 0; i < n_bufs; i++) {
@ -646,8 +646,8 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
free(galloc->hash_set.keys);
free(galloc->hash_values);
galloc->hash_set.size = hash_size;
galloc->hash_set.keys = calloc(sizeof(struct ggml_tensor *), hash_size);
galloc->hash_values = calloc(sizeof(struct hash_node), hash_size);
galloc->hash_set.keys = calloc(hash_size, sizeof(struct ggml_tensor *));
galloc->hash_values = calloc(hash_size, sizeof(struct hash_node));
GGML_ASSERT(galloc->hash_set.keys != NULL);
GGML_ASSERT(galloc->hash_values != NULL);
} else {
@ -667,7 +667,7 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
// set the node_allocs from the hash table
if (galloc->n_nodes < graph->n_nodes) {
free(galloc->node_allocs);
galloc->node_allocs = calloc(sizeof(struct node_alloc), graph->n_nodes);
galloc->node_allocs = calloc(graph->n_nodes, sizeof(struct node_alloc));
GGML_ASSERT(galloc->node_allocs != NULL);
}
galloc->n_nodes = graph->n_nodes;
@ -697,7 +697,7 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
}
if (galloc->n_leafs < graph->n_leafs) {
free(galloc->leaf_allocs);
galloc->leaf_allocs = calloc(sizeof(galloc->leaf_allocs[0]), graph->n_leafs);
galloc->leaf_allocs = calloc(graph->n_leafs, sizeof(galloc->leaf_allocs[0]));
GGML_ASSERT(galloc->leaf_allocs != NULL);
}
galloc->n_leafs = graph->n_leafs;

18
ggml-backend.c
View file

@ -1725,23 +1725,23 @@ ggml_backend_sched_t ggml_backend_sched_new(
GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU
struct ggml_backend_sched * sched = calloc(sizeof(struct ggml_backend_sched), 1);
struct ggml_backend_sched * sched = calloc(1, sizeof(struct ggml_backend_sched));
// initialize hash table
sched->hash_set = ggml_hash_set_new(graph_size);
sched->tensor_backend_id = calloc(sizeof(sched->tensor_backend_id[0]), sched->hash_set.size);
sched->tensor_copies = calloc(sizeof(sched->tensor_copies[0]), sched->hash_set.size);
sched->tensor_backend_id = calloc(sched->hash_set.size, sizeof(sched->tensor_backend_id[0]));
sched->tensor_copies = calloc(sched->hash_set.size, sizeof(sched->tensor_copies[0]));
const size_t nodes_size = graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2;
sched->node_backend_ids = calloc(sizeof(sched->node_backend_ids[0]), nodes_size);
sched->leaf_backend_ids = calloc(sizeof(sched->leaf_backend_ids[0]), nodes_size);
sched->node_backend_ids = calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
sched->leaf_backend_ids = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
sched->n_backends = n_backends;
sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
const int initial_splits_capacity = 16;
sched->splits = calloc(sizeof(sched->splits[0]), initial_splits_capacity);
sched->splits = calloc(initial_splits_capacity, sizeof(sched->splits[0]));
sched->splits_capacity = initial_splits_capacity;
for (int b = 0; b < n_backends; b++) {
@ -1972,10 +1972,10 @@ static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_te
struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) {
struct ggml_hash_set hash_set = {
/* .size = */ graph->visited_hash_table.size,
/* .keys = */ calloc(sizeof(hash_set.keys[0]), graph->visited_hash_table.size) // NOLINT
/* .keys = */ calloc(graph->visited_hash_table.size, sizeof(hash_set.keys[0])) // NOLINT
};
struct ggml_tensor ** node_copies = calloc(sizeof(node_copies[0]), hash_set.size); // NOLINT
bool * node_init = calloc(sizeof(node_init[0]), hash_set.size);
struct ggml_tensor ** node_copies = calloc(hash_set.size, sizeof(node_copies[0])); // NOLINT
bool * node_init = calloc(hash_set.size, sizeof(node_init[0]));
struct ggml_init_params params = {
/* .mem_size = */ ggml_tensor_overhead()*hash_set.size + ggml_graph_overhead_custom(graph->size, false),

260
ggml-impl.h
View file

@ -45,7 +45,7 @@ extern "C" {
// 16-bit float
// on Arm, we use __fp16
// on x86, we use uint16_t
#if defined(__ARM_NEON) && !defined(_MSC_VER)
#if defined(__ARM_NEON)
// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
//
@ -53,8 +53,262 @@ extern "C" {
//
#include <arm_neon.h>
#ifdef _MSC_VER
typedef uint16_t ggml_fp16_internal_t;
#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
#else
typedef __fp16 ggml_fp16_internal_t;
#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
#endif // _MSC_VER
#if !defined(__aarch64__)
// 32-bit ARM compatibility
// vaddvq_s16
// vpaddq_s16
// vpaddq_s32
// vaddvq_s32
// vaddvq_f32
// vmaxvq_f32
// vcvtnq_s32_f32
// vzip1_u8
// vzip2_u8
inline static int32_t vaddvq_s16(int16x8_t v) {
return
(int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) +
(int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) +
(int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) +
(int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7);
}
inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
return vcombine_s16(a0, b0);
}
inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
return vcombine_s32(a0, b0);
}
inline static int32_t vaddvq_s32(int32x4_t v) {
return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
}
inline static float vaddvq_f32(float32x4_t v) {
return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
}
inline static float vmaxvq_f32(float32x4_t v) {
return
MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
}
inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
int32x4_t res;
res[0] = roundf(vgetq_lane_f32(v, 0));
res[1] = roundf(vgetq_lane_f32(v, 1));
res[2] = roundf(vgetq_lane_f32(v, 2));
res[3] = roundf(vgetq_lane_f32(v, 3));
return res;
}
inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
uint8x8_t res;
res[0] = a[0]; res[1] = b[0];
res[2] = a[1]; res[3] = b[1];
res[4] = a[2]; res[5] = b[2];
res[6] = a[3]; res[7] = b[3];
return res;
}
inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
uint8x8_t res;
res[0] = a[4]; res[1] = b[4];
res[2] = a[5]; res[3] = b[5];
res[4] = a[6]; res[5] = b[6];
res[6] = a[7]; res[7] = b[7];
return res;
}
// vld1q_s16_x2
// vld1q_u8_x2
// vld1q_u8_x4
// vld1q_s8_x2
// vld1q_s8_x4
// TODO: double-check these work correctly
typedef struct ggml_int16x8x2_t {
int16x8_t val[2];
} ggml_int16x8x2_t;
inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
ggml_int16x8x2_t res;
res.val[0] = vld1q_s16(ptr + 0);
res.val[1] = vld1q_s16(ptr + 8);
return res;
}
typedef struct ggml_uint8x16x2_t {
uint8x16_t val[2];
} ggml_uint8x16x2_t;
inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
ggml_uint8x16x2_t res;
res.val[0] = vld1q_u8(ptr + 0);
res.val[1] = vld1q_u8(ptr + 16);
return res;
}
typedef struct ggml_uint8x16x4_t {
uint8x16_t val[4];
} ggml_uint8x16x4_t;
inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
ggml_uint8x16x4_t res;
res.val[0] = vld1q_u8(ptr + 0);
res.val[1] = vld1q_u8(ptr + 16);
res.val[2] = vld1q_u8(ptr + 32);
res.val[3] = vld1q_u8(ptr + 48);
return res;
}
typedef struct ggml_int8x16x2_t {
int8x16_t val[2];
} ggml_int8x16x2_t;
inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
ggml_int8x16x2_t res;
res.val[0] = vld1q_s8(ptr + 0);
res.val[1] = vld1q_s8(ptr + 16);
return res;
}
typedef struct ggml_int8x16x4_t {
int8x16_t val[4];
} ggml_int8x16x4_t;
inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
ggml_int8x16x4_t res;
res.val[0] = vld1q_s8(ptr + 0);
res.val[1] = vld1q_s8(ptr + 16);
res.val[2] = vld1q_s8(ptr + 32);
res.val[3] = vld1q_s8(ptr + 48);
return res;
}
// NOTE: not tested
inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
int8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
// NOTE: not tested
inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
uint8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
#else
#define ggml_int16x8x2_t int16x8x2_t
#define ggml_uint8x16x2_t uint8x16x2_t
#define ggml_uint8x16x4_t uint8x16x4_t
#define ggml_int8x16x2_t int8x16x2_t
#define ggml_int8x16x4_t int8x16x4_t
#define ggml_vld1q_s16_x2 vld1q_s16_x2
#define ggml_vld1q_u8_x2 vld1q_u8_x2
#define ggml_vld1q_u8_x4 vld1q_u8_x4
#define ggml_vld1q_s8_x2 vld1q_s8_x2
#define ggml_vld1q_s8_x4 vld1q_s8_x4
#define ggml_vqtbl1q_s8 vqtbl1q_s8
#define ggml_vqtbl1q_u8 vqtbl1q_u8
#endif // !defined(__aarch64__)
#if !defined(__ARM_FEATURE_DOTPROD)
inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
}
#else
#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
#endif // !defined(__ARM_FEATURE_DOTPROD)
#endif // defined(__ARM_NEON)
#if defined(__ARM_NEON) && !defined(__MSC_VER)
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
@ -75,8 +329,6 @@ static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
#else
typedef uint16_t ggml_fp16_internal_t;
#ifdef __wasm_simd128__
#include <wasm_simd128.h>
#else
@ -221,7 +473,7 @@ static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
#endif // __F16C__
#endif // __ARM_NEON
#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
// precomputed f32 table for f16 (256 KB)
// defined in ggml.c, initialized in ggml_init()

287
ggml-quants.c
View file

@ -14,41 +14,6 @@
#include <stdlib.h> // for qsort
#include <stdio.h> // for GGML_ASSERT
#ifdef __ARM_NEON
// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
//
// $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
//
#include <arm_neon.h>
#else
#ifdef __wasm_simd128__
#include <wasm_simd128.h>
#else
#if defined(__POWER9_VECTOR__) || defined(__powerpc64__)
#include <altivec.h>
#undef bool
#define bool _Bool
#else
#if defined(_MSC_VER) || defined(__MINGW32__)
#include <intrin.h>
#else
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__)
#if !defined(__riscv)
#include <immintrin.h>
#endif
#endif
#endif
#endif
#endif
#endif
#ifdef __riscv_v_intrinsic
#include <riscv_vector.h>
#endif
#undef MIN
#undef MAX
@ -276,258 +241,6 @@ static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128
#endif // __AVX__ || __AVX2__ || __AVX512F__
#endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
#if defined(__ARM_NEON)
#ifdef _MSC_VER
#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
#else
#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
#endif
#if !defined(__aarch64__)
// 64-bit compatibility
// vaddvq_s16
// vpaddq_s16
// vpaddq_s32
// vaddvq_s32
// vaddvq_f32
// vmaxvq_f32
// vcvtnq_s32_f32
// vzip1_u8
// vzip2_u8
inline static int32_t vaddvq_s16(int16x8_t v) {
return
(int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) +
(int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) +
(int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) +
(int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7);
}
inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
return vcombine_s16(a0, b0);
}
inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
return vcombine_s32(a0, b0);
}
inline static int32_t vaddvq_s32(int32x4_t v) {
return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
}
inline static float vaddvq_f32(float32x4_t v) {
return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
}
inline static float vmaxvq_f32(float32x4_t v) {
return
MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
}
inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
int32x4_t res;
res[0] = roundf(vgetq_lane_f32(v, 0));
res[1] = roundf(vgetq_lane_f32(v, 1));
res[2] = roundf(vgetq_lane_f32(v, 2));
res[3] = roundf(vgetq_lane_f32(v, 3));
return res;
}
inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
uint8x8_t res;
res[0] = a[0]; res[1] = b[0];
res[2] = a[1]; res[3] = b[1];
res[4] = a[2]; res[5] = b[2];
res[6] = a[3]; res[7] = b[3];
return res;
}
inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
uint8x8_t res;
res[0] = a[4]; res[1] = b[4];
res[2] = a[5]; res[3] = b[5];
res[4] = a[6]; res[5] = b[6];
res[6] = a[7]; res[7] = b[7];
return res;
}
// vld1q_s16_x2
// vld1q_u8_x2
// vld1q_u8_x4
// vld1q_s8_x2
// vld1q_s8_x4
// TODO: double-check these work correctly
typedef struct ggml_int16x8x2_t {
int16x8_t val[2];
} ggml_int16x8x2_t;
inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
ggml_int16x8x2_t res;
res.val[0] = vld1q_s16(ptr + 0);
res.val[1] = vld1q_s16(ptr + 8);
return res;
}
typedef struct ggml_uint8x16x2_t {
uint8x16_t val[2];
} ggml_uint8x16x2_t;
inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
ggml_uint8x16x2_t res;
res.val[0] = vld1q_u8(ptr + 0);
res.val[1] = vld1q_u8(ptr + 16);
return res;
}
typedef struct ggml_uint8x16x4_t {
uint8x16_t val[4];
} ggml_uint8x16x4_t;
inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
ggml_uint8x16x4_t res;
res.val[0] = vld1q_u8(ptr + 0);
res.val[1] = vld1q_u8(ptr + 16);
res.val[2] = vld1q_u8(ptr + 32);
res.val[3] = vld1q_u8(ptr + 48);
return res;
}
typedef struct ggml_int8x16x2_t {
int8x16_t val[2];
} ggml_int8x16x2_t;
inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
ggml_int8x16x2_t res;
res.val[0] = vld1q_s8(ptr + 0);
res.val[1] = vld1q_s8(ptr + 16);
return res;
}
typedef struct ggml_int8x16x4_t {
int8x16_t val[4];
} ggml_int8x16x4_t;
inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
ggml_int8x16x4_t res;
res.val[0] = vld1q_s8(ptr + 0);
res.val[1] = vld1q_s8(ptr + 16);
res.val[2] = vld1q_s8(ptr + 32);
res.val[3] = vld1q_s8(ptr + 48);
return res;
}
// NOTE: not tested
inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
int8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
// NOTE: not tested
inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
uint8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
#else
#define ggml_int16x8x2_t int16x8x2_t
#define ggml_uint8x16x2_t uint8x16x2_t
#define ggml_uint8x16x4_t uint8x16x4_t
#define ggml_int8x16x2_t int8x16x2_t
#define ggml_int8x16x4_t int8x16x4_t
#define ggml_vld1q_s16_x2 vld1q_s16_x2
#define ggml_vld1q_u8_x2 vld1q_u8_x2
#define ggml_vld1q_u8_x4 vld1q_u8_x4
#define ggml_vld1q_s8_x2 vld1q_s8_x2
#define ggml_vld1q_s8_x4 vld1q_s8_x4
#define ggml_vqtbl1q_s8 vqtbl1q_s8
#define ggml_vqtbl1q_u8 vqtbl1q_u8
#endif
#if !defined(__ARM_FEATURE_DOTPROD)
inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
}
#else
#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
#endif
#endif
#if defined(__ARM_NEON) || defined(__wasm_simd128__)
#define B1(c,s,n) 0x ## n ## c , 0x ## n ## s
#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)

8
ggml.c
View file

@ -10825,7 +10825,7 @@ static void ggml_compute_forward_mul_mat(
#endif
#if GGML_USE_LLAMAFILE
if (nb10 == ggml_type_size(src1->type)) {
if (src1_cont) {
for (int64_t i13 = 0; i13 < ne13; i13++)
for (int64_t i12 = 0; i12 < ne12; i12++)
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
@ -10878,15 +10878,13 @@ UseGgmlGemm1:;
const size_t row_size = ggml_row_size(vec_dot_type, ne10);
#if GGML_USE_LLAMAFILE
if (nb10 == ggml_type_size(src1->type) || src1->type != vec_dot_type) {
if (src1->type != vec_dot_type) {
for (int64_t i13 = 0; i13 < ne13; i13++)
for (int64_t i12 = 0; i12 < ne12; i12++)
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
(const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
nb01/ggml_type_size(src0->type),
(const char *)wdata + ggml_row_size(vec_dot_type,
nb12/ggml_type_size(src1->type)*i12 +
nb13/ggml_type_size(src1->type)*i13),
(const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size,
row_size/ggml_type_size(vec_dot_type),
(char *)dst->data + i12*nb2 + i13*nb3,
nb1/ggml_type_size(dst->type),

16
gguf-py/gguf/constants.py
View file

@ -124,6 +124,7 @@ class MODEL_ARCH(IntEnum):
QWEN2 = auto()
QWEN2MOE = auto()
PHI2 = auto()
PHI3 = auto()
PLAMO = auto()
CODESHELL = auto()
ORION = auto()
@ -200,6 +201,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.QWEN2: "qwen2",
MODEL_ARCH.QWEN2MOE: "qwen2moe",
MODEL_ARCH.PHI2: "phi2",
MODEL_ARCH.PHI3: "phi3",
MODEL_ARCH.PLAMO: "plamo",
MODEL_ARCH.CODESHELL: "codeshell",
MODEL_ARCH.ORION: "orion",
@ -550,6 +552,20 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.PHI3: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.CODESHELL: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.POS_EMBD,

2
gguf-py/gguf/tensor_mapping.py
View file

@ -117,6 +117,7 @@ class TensorNameMap:
"h.{bid}.attn.c_attn", # gpt2
"transformer.h.{bid}.mixer.Wqkv", # phi2
"encoder.layers.{bid}.attn.Wqkv", # nomic-bert
"model.layers.{bid}.self_attn.qkv_proj" # phi3
),
# Attention query
@ -234,6 +235,7 @@ class TensorNameMap:
"h.{bid}.mlp.c_fc", # gpt2
"transformer.h.{bid}.mlp.fc1", # phi2
"model.layers.{bid}.mlp.fc1", # phi2
"model.layers.{bid}.mlp.gate_up_proj", # phi3
"model.layers.layers.{bid}.mlp.up_proj", # plamo
"model.layers.{bid}.feed_forward.w3", # internlm2
"encoder.layers.{bid}.mlp.fc11", # nomic-bert

216
llama.cpp
View file

@ -211,6 +211,7 @@ enum llm_arch {
LLM_ARCH_QWEN2,
LLM_ARCH_QWEN2MOE,
LLM_ARCH_PHI2,
LLM_ARCH_PHI3,
LLM_ARCH_PLAMO,
LLM_ARCH_CODESHELL,
LLM_ARCH_ORION,
@ -246,6 +247,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_QWEN2, "qwen2" },
{ LLM_ARCH_QWEN2MOE, "qwen2moe" },
{ LLM_ARCH_PHI2, "phi2" },
{ LLM_ARCH_PHI3, "phi3" },
{ LLM_ARCH_PLAMO, "plamo" },
{ LLM_ARCH_CODESHELL, "codeshell" },
{ LLM_ARCH_ORION, "orion" },
@ -793,6 +795,23 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
LLM_ARCH_PHI3,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
LLM_ARCH_PLAMO,
{
@ -3955,6 +3974,16 @@ static void llm_load_hparams(
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
switch (hparams.n_layer) {
case 24: model.type = e_model::MODEL_1B; break;
case 32: model.type = e_model::MODEL_3B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_PHI3:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) {
case 24: model.type = e_model::MODEL_1B; break;
case 32: model.type = e_model::MODEL_3B; break;
@ -4340,7 +4369,7 @@ static void llm_load_vocab(
}
}
// find EOT token: "<|eot_id|>", "<|im_emd|>", "<end_of_turn>", etc.
// find EOT token: "<|eot_id|>", "<|im_end|>", "<end_of_turn>", etc.
//
// TODO: convert scripts should provide this token through the KV metadata LLAMA_KV_TOKENIZER_EOT_ID
// for now, we apply this workaround to find the EOT token based on its text
@ -4351,7 +4380,8 @@ static void llm_load_vocab(
// need to fix convert script
//vocab.id_to_token[t.second].type == LLAMA_TOKEN_TYPE_CONTROL &&
(t.first == "<|eot_id|>" ||
t.first == "<|im_emd|>" ||
t.first == "<|im_end|>" ||
t.first == "<|end|>" ||
t.first == "<end_of_turn>"
)
) {
@ -5375,6 +5405,33 @@ static bool llm_load_tensors(
layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff});
}
} break;
case LLM_ARCH_PHI3:
{
model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab });
// output
{
model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd });
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab });
}
for (int i = 0; i < n_layer; ++i) {
ggml_context* ctx_layer = ctx_for_layer(i);
ggml_context* ctx_split = ctx_for_layer_split(i);
auto& layer = model.layers[i];
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.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_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd });
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, 2 * n_ff });
}
} break;
case LLM_ARCH_PLAMO:
{
model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
@ -6326,7 +6383,7 @@ static struct ggml_tensor * llm_build_kqv(
struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
cb(kq, "kq", il);
if (model.arch == LLM_ARCH_PHI2) {
if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3) {
// for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
// ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
@ -8967,12 +9024,140 @@ struct llm_build_context {
cur = ggml_add(ctx0, cur, model.output_b);
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}
struct ggml_cgraph * build_phi3() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
// inp_pos - contains the positions
struct ggml_tensor * inp_pos = build_inp_pos();
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
for (int il = 0; il < n_layer; ++il) {
auto residual = inpL;
// self-attention
{
struct ggml_tensor* attn_norm_output = llm_build_norm(ctx0, inpL, hparams,
model.layers[il].attn_norm,
NULL,
LLM_NORM_RMS, cb, il);
cb(attn_norm_output, "attn_norm", il);
struct ggml_tensor * Qcur = nullptr;
struct ggml_tensor * Kcur = nullptr;
struct ggml_tensor * Vcur = nullptr;
if (model.layers[il].wqkv) {
cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, attn_norm_output);
cb(cur, "wqkv", il);
Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0 * sizeof(float) * (n_embd)));
Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd)));
Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa)));
}
else {
Qcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, attn_norm_output), model.layers[il].bq);
Kcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, attn_norm_output), model.layers[il].bk);
Vcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, attn_norm_output), model.layers[il].bv);
}
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
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);
Qcur = ggml_rope_custom(
ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head)));
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, NULL,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f, cb, il);
}
if (il == n_layer - 1) {
// skip computing output for unused tokens
struct ggml_tensor* inp_out_ids = build_inp_out_ids();
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
residual = ggml_get_rows(ctx0, residual, inp_out_ids);
}
cur = ggml_add(ctx0, cur, residual);
residual = cur;
cur = llm_build_norm(ctx0, cur, hparams,
model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, cb, il);
cb(cur, "ffn_norm", il);
// FF
// special-case: the up and gate tensors are merged into a single tensor
// TOOD: support into llm_build_ffn
{
struct ggml_tensor* up = ggml_mul_mat(ctx0, model.layers[il].ffn_up, cur);
cb(up, "ffn_up", il);
auto g = ggml_cont(ctx0, ggml_view_2d(ctx0, up, up->ne[0] / 2, up->ne[1], ggml_row_size(up->type, up->ne[0]), 0));
auto y = ggml_cont(ctx0, ggml_view_2d(ctx0, up, up->ne[0] / 2, up->ne[1], ggml_row_size(up->type, up->ne[0]), up->nb[1] / 2));
y = ggml_mul(ctx0, y, ggml_silu(ctx0, g));
cb(y, "ffn_gate", il);
auto down = ggml_mul_mat(ctx0, model.layers[il].ffn_down, y);
cb(down, "ffn_down", il);
cur = down;
cb(cur, "ffn_out", il);
}
cur = ggml_add(ctx0, residual, cur);
cb(cur, "l_out", il);
inpL = cur;
}
cur = llm_build_norm(ctx0, inpL, hparams,
model.output_norm,
NULL,
LLM_NORM_RMS, cb, -1);
cb(cur, "result_norm", -1);
cur = ggml_mul_mat(ctx0, model.output, cur);
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}
struct ggml_cgraph * build_plamo() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
@ -10474,6 +10659,10 @@ static struct ggml_cgraph * llama_build_graph(
{
result = llm.build_phi2();
} break;
case LLM_ARCH_PHI3:
{
result = llm.build_phi3();
} break;
case LLM_ARCH_PLAMO:
{
result = llm.build_plamo();
@ -13478,7 +13667,7 @@ llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_da
return result;
}
llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates) {
llama_token llama_sample_token_with_rng(struct llama_context * ctx, llama_token_data_array * candidates, std::mt19937 & rng) {
GGML_ASSERT(ctx);
const int64_t t_start_sample_us = ggml_time_us();
@ -13491,7 +13680,6 @@ llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_arra
}
std::discrete_distribution<> dist(probs.begin(), probs.end());
auto & rng = ctx->rng;
int idx = dist(rng);
llama_token result = candidates->data[idx].id;
@ -13501,6 +13689,10 @@ llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_arra
return result;
}
llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates) {
return llama_sample_token_with_rng(ctx, candidates, ctx->rng);
}
void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token) {
const int64_t t_start_sample_us = ggml_time_us();
@ -15393,6 +15585,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
case LLM_ARCH_QWEN2:
case LLM_ARCH_QWEN2MOE:
case LLM_ARCH_PHI2:
case LLM_ARCH_PHI3:
case LLM_ARCH_GEMMA:
case LLM_ARCH_STARCODER2:
return LLAMA_ROPE_TYPE_NEOX;
@ -15406,6 +15599,10 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
return LLAMA_ROPE_TYPE_NONE;
}
enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx) {
return ctx->cparams.pooling_type;
}
int32_t llama_n_vocab(const struct llama_model * model) {
return model->hparams.n_vocab;
}
@ -17098,6 +17295,15 @@ LLAMA_API int32_t llama_chat_get_model_template(
if (model_template.empty()) {
model_template = get_meta(default_meta);
}
} else if (tmpl == "phi3" || (tmpl.find("<|assistant|>") != std::string::npos && tmpl.find("<|end|>") != std::string::npos )) {
// Phi 3
for (auto message : chat) {
std::string role(message->role);
ss << "<|" << role << "|>\n" << trim(message->content) << "<|end|>\n";
}
if (add_ass) {
ss << "<|assistant|>\n";
}
} else {
// default template
model_template = get_meta(default_meta);

15
llama.h
View file

@ -408,8 +408,10 @@ extern "C" {
LLAMA_API uint32_t llama_n_ubatch (const struct llama_context * ctx);
LLAMA_API uint32_t llama_n_seq_max (const struct llama_context * ctx);
LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model);
LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model);
LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx);
LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model);
LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model);
LLAMA_API int32_t llama_n_vocab (const struct llama_model * model);
LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
@ -1057,7 +1059,7 @@ extern "C" {
struct llama_context * ctx,
llama_token_data_array * candidates);
/// @details Randomly selects a token from the candidates based on their probabilities.
/// @details Randomly selects a token from the candidates based on their probabilities using the RNG of ctx.
LLAMA_API llama_token llama_sample_token(
struct llama_context * ctx,
llama_token_data_array * candidates);
@ -1144,8 +1146,9 @@ extern "C" {
// Internal API to be implemented by llama.cpp and used by tests/benchmarks only
#ifdef LLAMA_API_INTERNAL
#include <vector>
#include <random>
#include <string>
#include <vector>
struct ggml_tensor;
@ -1182,6 +1185,10 @@ std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
const std::string & src,
llama_partial_utf8 partial_start);
// Randomly selects a token from the candidates based on their probabilities using given std::mt19937.
// This is a temporary workaround in order to fix race conditions when sampling with multiple sequences.
llama_token llama_sample_token_with_rng(struct llama_context * ctx, llama_token_data_array * candidates, std::mt19937 & rng);
#endif // LLAMA_API_INTERNAL
#endif // LLAMA_H

945
sgemm.cpp
View file

File diff suppressed because it is too large Load diff

4
tests/test-chat-template.cpp
View file

@ -49,6 +49,8 @@ int main(void) {
"{{ bos_token }}{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif false == true %}{% set loop_messages = messages %}{% set system_message = 'You are Command-R, a brilliant, sophisticated, AI-assistant trained to assist human users by providing thorough responses. You are trained by Cohere.' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% if system_message != false %}{{ '<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>' + system_message + '<|END_OF_TURN_TOKEN|>' }}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% set content = message['content'] %}{% if message['role'] == 'user' %}{{ '<|START_OF_TURN_TOKEN|><|USER_TOKEN|>' + content.strip() + '<|END_OF_TURN_TOKEN|>' }}{% elif message['role'] == 'assistant' %}{{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>' + content.strip() + '<|END_OF_TURN_TOKEN|>' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>' }}{% endif %}",
// Llama-3
"{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}",
// Phi-3
"{{ bos_token }}{% for message in messages %}{{'<|' + message['role'] + '|>' + ' ' + message['content'] + '<|end|> ' }}{% endfor %}{% if add_generation_prompt %}{{ '<|assistant|> ' }}{% else %}{{ eos_token }}{% endif %}"
};
std::vector<std::string> expected_output = {
// teknium/OpenHermes-2.5-Mistral-7B
@ -77,6 +79,8 @@ int main(void) {
"<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>You are a helpful assistant<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Hello<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>Hi there<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Who are you<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>I am an assistant<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Another question<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>",
// Llama 3
"<|start_header_id|>system<|end_header_id|>\n\nYou are a helpful assistant<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHello<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nHi there<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWho are you<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nI am an assistant<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nAnother question<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n",
// Phi 3
"<|system|>\nYou are a helpful assistant<|end|>\n<|user|>\nHello<|end|>\n<|assistant|>\nHi there<|end|>\n<|user|>\nWho are you<|end|>\n<|assistant|>\nI am an assistant<|end|>\n<|user|>\nAnother question<|end|>\n<|assistant|>\n",
};
std::vector<char> formatted_chat(1024);
int32_t res;