Commit 97ef6ff8 authored by xuxzh1's avatar xuxzh1 🎱
Browse files

update

parent 4cc1a614
Pipeline #2023 canceled with stages
......@@ -77,6 +77,35 @@ Feature: Parallel
| disabled | 128 |
| enabled | 64 |
Scenario Outline: Multi users with number of prompts exceeding number of slots
Given a system prompt You are a writer.
And a model tinyllama-2
Given a prompt:
"""
Write a very long book.
"""
And a prompt:
"""
Write another a poem.
"""
And a prompt:
"""
What is LLM?
"""
And a prompt:
"""
The sky is blue and I love it.
"""
And <n_predict> max tokens to predict
And streaming is <streaming>
Given concurrent OAI completions requests
Then the server is busy
Then the server is idle
Then all prompts are predicted with <n_predict> tokens
Examples:
| streaming | n_predict |
| disabled | 128 |
| enabled | 64 |
Scenario: Multi users with total number of tokens to predict exceeds the KV Cache size #3969
Given a prompt:
......
......@@ -15,6 +15,7 @@ Feature: Passkey / Self-extend with context shift
And <n_junk> as number of junk
And <n_predicted> server max tokens to predict
And 42 as seed
And 0.0 temperature
And <n_ctx> KV cache size
And 1 slots
And <n_ga> group attention factor to extend context size through self-extend
......@@ -22,7 +23,8 @@ Feature: Passkey / Self-extend with context shift
# Can be override with N_GPU_LAYERS
And <ngl> GPU offloaded layers
Then the server is starting
Then the server is healthy
# Higher timeout because the model may need to be downloaded from the internet
Then the server is healthy with timeout 120 seconds
Given available models
Then model 0 is trained on <n_ctx_train> tokens context
Given a prefix prompt:
......
@llama.cpp
@rerank
Feature: llama.cpp server
Background: Server startup
Given a server listening on localhost:8080
And a model url https://huggingface.co/ggml-org/models/resolve/main/jina-reranker-v1-tiny-en/ggml-model-f16.gguf
And a model file jina-reranker-v1-tiny-en.gguf
And a model alias jina-reranker-v1-tiny-en
And 42 as server seed
And 2 slots
And 512 as batch size
And 512 as ubatch size
And 512 KV cache size
And enable reranking endpoint
Then the server is starting
Then the server is healthy
Scenario: Rerank
Given a rerank query:
"""
Machine learning is
"""
And a rerank document:
"""
A machine is a physical system that uses power to apply forces and control movement to perform an action. The term is commonly applied to artificial devices, such as those employing engines or motors, but also to natural biological macromolecules, such as molecular machines.
"""
And a rerank document:
"""
Learning is the process of acquiring new understanding, knowledge, behaviors, skills, values, attitudes, and preferences. The ability to learn is possessed by humans, non-human animals, and some machines; there is also evidence for some kind of learning in certain plants.
"""
And a rerank document:
"""
Machine learning is a field of study in artificial intelligence concerned with the development and study of statistical algorithms that can learn from data and generalize to unseen data, and thus perform tasks without explicit instructions.
"""
And a rerank document:
"""
Paris, capitale de la France, est une grande ville européenne et un centre mondial de l'art, de la mode, de la gastronomie et de la culture. Son paysage urbain du XIXe siècle est traversé par de larges boulevards et la Seine.
"""
When reranking request
Then reranking results are returned
Then reranking highest score is index 2 and lowest score is index 3
......@@ -5,7 +5,7 @@ Feature: Security
Background: Server startup with an api key defined
Given a server listening on localhost:8080
And a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
And a server api key llama.cpp
And a server api key THIS_IS_THE_KEY
Then the server is starting
Then the server is healthy
......@@ -16,11 +16,11 @@ Feature: Security
And a completion request with <api_error> api error
Examples: Prompts
| api_key | api_error |
| llama.cpp | no |
| llama.cpp | no |
| hackeme | raised |
| | raised |
| api_key | api_error |
| THIS_IS_THE_KEY | no |
| THIS_IS_THE_KEY | no |
| hackeme | raised |
| | raised |
Scenario Outline: OAI Compatibility
Given a system prompt test
......@@ -32,10 +32,10 @@ Feature: Security
Given an OAI compatible chat completions request with <api_error> api error
Examples: Prompts
| api_key | api_error |
| llama.cpp | no |
| llama.cpp | no |
| hackme | raised |
| api_key | api_error |
| THIS_IS_THE_KEY | no |
| THIS_IS_THE_KEY | no |
| hackme | raised |
Scenario Outline: OAI Compatibility (invalid response formats)
Given a system prompt test
......@@ -55,7 +55,7 @@ Feature: Security
Scenario Outline: CORS Options
Given a user api key llama.cpp
Given a user api key THIS_IS_THE_KEY
When an OPTIONS request is sent from <origin>
Then CORS header <cors_header> is set to <cors_header_value>
......@@ -64,5 +64,5 @@ Feature: Security
| localhost | Access-Control-Allow-Origin | localhost |
| web.mydomain.fr | Access-Control-Allow-Origin | web.mydomain.fr |
| origin | Access-Control-Allow-Credentials | true |
| web.mydomain.fr | Access-Control-Allow-Methods | POST |
| web.mydomain.fr | Access-Control-Allow-Methods | GET, POST |
| web.mydomain.fr | Access-Control-Allow-Headers | * |
......@@ -105,6 +105,14 @@ Feature: llama.cpp server
Given first token is removed
Then tokens can be detokenized
Scenario: Tokenize with pieces
When tokenizing with pieces:
"""
What is the capital of Germany?
"""
Then tokens are given with pieces
Scenario: Models available
Given available models
Then 1 models are supported
......
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
import asyncio
import json
import os
......@@ -23,6 +26,8 @@ from prometheus_client import parser
# pyright: reportRedeclaration=false
DEFAULT_TIMEOUT_SECONDS = aiohttp.ClientTimeout(total=600)
@step("a server listening on {server_fqdn}:{server_port}")
def step_server_config(context, server_fqdn: str, server_port: str):
context.server_fqdn = server_fqdn
......@@ -63,6 +68,7 @@ def step_server_config(context, server_fqdn: str, server_port: str):
context.server_api_key = None
context.server_continuous_batching = False
context.server_embeddings = False
context.server_reranking = False
context.server_metrics = False
context.server_process = None
context.seed = None
......@@ -72,11 +78,21 @@ def step_server_config(context, server_fqdn: str, server_port: str):
context.response_format = None
context.temperature = None
context.lora_file = None
context.disable_ctx_shift = False
# infill
context.infill_input_extra = None
context.infill_input_suffix = ''
context.infill_input_prefix = ''
context.tasks_result = []
context.concurrent_tasks = []
context.prompts = []
context.reranking_query = None
context.reranking_documents = []
context.reranking_results = None
@step('a model file {hf_file} from HF repo {hf_repo}')
def step_download_hf_model(context, hf_file: str, hf_repo: str):
......@@ -143,7 +159,7 @@ def step_n_slots(context, n_slots: int):
@step('{n_predict:d} server max tokens to predict')
def step_server_n_predict(context, n_predict: int):
context.n_server_predict = n_predict
context.n_server_predict = n_predict if n_predict > 0 else None
@step('{slot_save_path} as slot save path')
......@@ -166,15 +182,21 @@ def step_server_continuous_batching(context):
context.server_continuous_batching = True
@step('embeddings extraction')
@step('enable embeddings endpoint')
def step_server_embeddings(context):
context.server_embeddings = True
@step('enable reranking endpoint')
def step_server_reranking(context):
context.server_reranking = True
@step('prometheus compatible metrics exposed')
def step_server_metrics(context):
context.server_metrics = True
@step('disable context shifting')
def step_server_disable_ctx_shift(context):
context.disable_ctx_shift = True
@step("the server is starting")
def step_start_server(context):
......@@ -200,48 +222,51 @@ def step_start_server(context):
time.sleep(0.1)
@step("the server is {expecting_status}")
@async_run_until_complete
async def step_wait_for_the_server_to_be_started(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str):
async def wait_for_server_status_with_timeout(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str, timeout: int):
match expecting_status:
case 'healthy':
await wait_for_health_status(context, context.base_url, 200, 'ok',
timeout=30)
await wait_for_slots_status(context, context.base_url, 200,
timeout=timeout)
case 'ready' | 'idle':
await wait_for_health_status(context, context.base_url, 200, 'ok',
timeout=30,
params={'fail_on_no_slot': 0, 'include_slots': 0},
slots_idle=context.n_slots,
slots_processing=0,
expected_slots=[{'id': slot_id, 'state': 0}
for slot_id in
range(context.n_slots if context.n_slots else 1)])
await wait_for_slots_status(context, context.base_url, 200,
timeout=timeout,
params={'fail_on_no_slot': 1},
slots_idle=context.n_slots,
slots_processing=0)
case 'busy':
await wait_for_health_status(context, context.base_url, 503,
'no slot available',
params={'fail_on_no_slot': 0, 'include_slots': 0},
slots_idle=0,
slots_processing=context.n_slots,
expected_slots=[{'id': slot_id, 'state': 1}
for slot_id in
range(context.n_slots if context.n_slots else 1)])
await wait_for_slots_status(context, context.base_url, 503,
params={'fail_on_no_slot': 1},
slots_idle=0,
slots_processing=context.n_slots)
case _:
assert False, "unknown status"
@step("the server is {expecting_status} with timeout {timeout:d} seconds")
@async_run_until_complete
async def step_wait_for_server_status_with_timeout(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str, timeout: int):
await wait_for_server_status_with_timeout(context, expecting_status, timeout)
@step("the server is {expecting_status}")
@async_run_until_complete
async def step_wait_for_server_status(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str):
await wait_for_server_status_with_timeout(context, expecting_status, 30)
@step('all slots are {expected_slot_status_string}')
@async_run_until_complete
async def step_all_slots_status(context, expected_slot_status_string: Literal['idle', 'busy'] | str):
match expected_slot_status_string:
case 'idle':
expected_slot_status = 0
expected_slot_status = False
case 'busy':
expected_slot_status = 1
expected_slot_status = True
case _:
assert False, "unknown status"
expected_slots = [{'id': slot_id, 'state': expected_slot_status}
expected_slots = [{'id': slot_id, 'is_processing': expected_slot_status}
for slot_id in range(context.n_slots)]
await request_slots_status(context, expected_slots)
......@@ -249,7 +274,7 @@ async def step_all_slots_status(context, expected_slot_status_string: Literal['i
@step('a completion request with {api_error} api error')
@async_run_until_complete
async def step_request_completion(context, api_error: Literal['raised'] | str):
expect_api_error = api_error == 'raised'
expect_api_error = api_error == 'raised' or api_error != 'no'
seeds = await completions_seed(context, num_seeds=1)
completion = await request_completion(context.prompts.pop(),
seeds[0] if seeds is not None else seeds,
......@@ -264,8 +289,33 @@ async def step_request_completion(context, api_error: Literal['raised'] | str):
context.tasks_result.append(completion)
if context.debug:
print(f"Completion response: {completion}")
if expect_api_error:
if api_error == 'raised':
assert completion == 401, f"completion must be an 401 status code: {completion}"
elif api_error.isdigit():
api_error_code = int(api_error)
assert completion == api_error_code, f"completion must be an {api_error_code} status code: {completion}"
@step('an infill request with {api_error} api error')
@async_run_until_complete
async def step_request_completion(context, api_error: Literal['raised'] | str):
if api_error != 'no':
raise ValueError(f'api_error={api_error} is not yet implemented')
payload = {
"prompt": context.prompts[0],
"input_suffix": context.infill_input_suffix,
"input_prefix": context.infill_input_prefix,
"n_predict": context.n_predict,
"seed": context.seed,
"temperature": context.temperature,
}
if context.infill_input_extra is not None:
payload['input_extra'] = context.infill_input_extra
async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
async with session.post(f'{context.base_url}/infill',
json=payload) as response:
assert response.status == 200
context.tasks_result = [await response.json()]
@step('{predicted_n:d} tokens are predicted matching {re_content}')
......@@ -437,6 +487,14 @@ def step_impl(context, n_ga_w):
def step_prompt_passkey(context):
context.prompt_passkey = context_text(context)
@step('a rerank query')
def step_set_rerank_query(context):
context.reranking_query = context_text(context)
context.reranking_documents = []
@step('a rerank document')
def step_set_rerank_document(context):
context.reranking_documents.append(context_text(context))
@step('{n_prompts:d} fixed prompts')
def step_fixed_prompts(context, n_prompts):
......@@ -508,6 +566,25 @@ def step_a_prompt_prompt(context, prompt):
context.n_prompts = len(context.prompts)
# TODO: allow this to be repeated
@step('an infill input extra {filename} {text}')
def step_infill_input_extra(context, filename, text):
if filename == 'none':
context.infill_input_extra = None
else:
context.infill_input_extra = [{'filename': filename, 'text': text}]
@step('an infill input suffix {text}')
def step_infill_input_suffix(context, text):
context.infill_input_suffix = text
@step('an infill input prefix {text}')
def step_infill_input_prefix(context, text):
context.infill_input_prefix = text
@step('{num_prompts:d} prompts {prompt} with seed {seed:d}')
def step_many_prompts(context, num_prompts, prompt, seed):
if context.seed is None:
......@@ -604,6 +681,22 @@ async def step_compute_embedding(context):
context.embeddings = await request_embedding(context_text(context), None, base_url=context.base_url)
@step('reranking request')
@async_run_until_complete
async def step_compute_reranking(context):
async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
async with session.post(f'{context.base_url}/reranking',
json={
"query": context.reranking_query,
"documents": context.reranking_documents,
}) as response:
if response.status == 200:
response_json = await response.json()
context.reranking_results = response_json['results']
else:
context.reranking_results = response.status
@step('all embeddings are the same')
@async_run_until_complete
async def step_all_embeddings_are_the_same(context):
......@@ -637,6 +730,9 @@ def step_assert_embeddings(context):
for embedding in context.embeddings:
assert_embeddings(embedding)
@step('embeddings request with {api_error_code:d} api error')
def step_assert_embeddings(context, api_error_code: int):
assert context.embeddings == api_error_code, f"embeddings request must return code {api_error_code}, but got {context.embeddings}"
@step('an OAI compatible embeddings computation request for')
@async_run_until_complete
......@@ -686,17 +782,61 @@ async def all_embeddings_are_generated(context):
for i in range(n_embedding_requests):
assert_embeddings(context.tasks_result.pop().pop())
@step('reranking results are returned')
def reranking_results_are_returned(context):
assert len(context.reranking_results) == len(context.reranking_documents)
@step('reranking highest score is index {idx_high:d} and lowest score is index {idx_low:d}')
def reranking_results_are_returned(context, idx_high: int, idx_low: int):
max_score, max_idx = 0, 0
min_score, min_idx = 0, 0
for res in context.reranking_results:
if max_score < res['relevance_score']:
max_score = res['relevance_score']
max_idx = res['index']
if min_score > res['relevance_score']:
min_score = res['relevance_score']
min_idx = res['index']
print(context.reranking_results)
assert max_idx == idx_high
assert min_idx == idx_low
@step('adding special tokens')
def step_tokenize_set_add_special(context):
context.tokenize_add_special = True
@step("tokenizing with pieces")
@async_run_until_complete
async def step_tokenize_with_pieces(context):
context.tokenized_text = context_text(context)
async with aiohttp.ClientSession() as session:
tokenize_args = {"content": context.tokenized_text, "with_pieces": True}
if getattr(context, "tokenize_add_special", None) is not None:
tokenize_args["add_special"] = context.tokenize_add_special
async with session.post(
f"{context.base_url}/tokenize", json=tokenize_args
) as response:
assert response.status == 200
tokenize_json = await response.json()
context.tokens_with_pieces = tokenize_json["tokens"]
@step("tokens are given with pieces")
@async_run_until_complete
async def step_tokenize_with_pieces(context):
# Verify that the response contains both token IDs and pieces
assert all(
"id" in token and "piece" in token for token in context.tokens_with_pieces
)
@step('tokenizing')
@async_run_until_complete
async def step_tokenize(context):
context.tokenized_text = context_text(context)
async with aiohttp.ClientSession() as session:
async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
tokenize_args = {
"content": context.tokenized_text,
}
......@@ -713,7 +853,7 @@ async def step_tokenize(context):
@async_run_until_complete
async def step_detokenize(context):
assert len(context.tokens) > 0
async with aiohttp.ClientSession() as session:
async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
async with session.post(f'{context.base_url}/detokenize',
json={
"tokens": context.tokens,
......@@ -742,7 +882,7 @@ def step_strings_for_tokenization(context):
@step('an OPTIONS request is sent from {origin}')
@async_run_until_complete
async def step_options_request(context, origin):
async with aiohttp.ClientSession() as session:
async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
headers = {'Authorization': f'Bearer {context.user_api_key}', 'Origin': origin}
async with session.options(f'{context.base_url}/v1/chat/completions',
headers=headers) as response:
......@@ -758,7 +898,7 @@ def step_check_options_header_value(context, cors_header, cors_header_value):
@step('prometheus metrics are exposed')
@async_run_until_complete
async def step_prometheus_metrics_exported(context):
async with aiohttp.ClientSession() as session:
async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
async with await session.get(f'{context.base_url}/metrics') as metrics_response:
assert metrics_response.status == 200
assert metrics_response.headers['Content-Type'] == "text/plain; version=0.0.4"
......@@ -825,13 +965,13 @@ async def concurrent_requests(context, f_completion, *args, **kwargs):
for prompt_no in range(context.n_prompts):
shifted_args = [context.prompts.pop(), seeds[prompt_no], *args]
context.concurrent_tasks.append(asyncio.create_task(f_completion(*shifted_args, **kwargs)))
await asyncio.sleep(0.1)
await asyncio.sleep(0.01)
@step('the slot {slot_id:d} is saved with filename "{filename}"')
@async_run_until_complete
async def step_save_slot(context, slot_id, filename):
async with aiohttp.ClientSession() as session:
async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
async with session.post(f'{context.base_url}/slots/{slot_id}?action=save',
json={"filename": filename},
headers={"Content-Type": "application/json"}) as response:
......@@ -841,7 +981,7 @@ async def step_save_slot(context, slot_id, filename):
@step('the slot {slot_id:d} is restored with filename "{filename}"')
@async_run_until_complete
async def step_restore_slot(context, slot_id, filename):
async with aiohttp.ClientSession() as session:
async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
async with session.post(f'{context.base_url}/slots/{slot_id}?action=restore',
json={"filename": filename},
headers={"Content-Type": "application/json"}) as response:
......@@ -851,7 +991,7 @@ async def step_restore_slot(context, slot_id, filename):
@step('the slot {slot_id:d} is erased')
@async_run_until_complete
async def step_erase_slot(context, slot_id):
async with aiohttp.ClientSession() as session:
async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
async with session.post(f'{context.base_url}/slots/{slot_id}?action=erase',
headers={"Content-Type": "application/json"}) as response:
context.response = response
......@@ -860,7 +1000,7 @@ async def step_erase_slot(context, slot_id):
@step('switch {on_or_off} lora adapter {lora_id:d}')
@async_run_until_complete
async def toggle_lora_adapter(context, on_or_off: str, lora_id: int):
async with aiohttp.ClientSession() as session:
async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
async with session.post(f'{context.base_url}/lora-adapters',
json=[{'id': lora_id, 'scale': 1 if on_or_off == 'on' else 0}],
headers={"Content-Type": "application/json"}) as response:
......@@ -896,7 +1036,7 @@ async def request_completion(prompt,
print(f"Set user_api_key: {user_api_key}")
headers['Authorization'] = f'Bearer {user_api_key}'
async with aiohttp.ClientSession() as session:
async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
async with session.post(f'{base_url}/completion',
json={
"input_prefix": prompt_prefix,
......@@ -909,8 +1049,7 @@ async def request_completion(prompt,
"temperature": temperature if temperature is not None else 0.8,
"n_probs": 2,
},
headers=headers,
timeout=3600) as response:
headers=headers) as response:
if expect_api_error is None or not expect_api_error:
assert response.status == 200
assert response.headers['Access-Control-Allow-Origin'] == origin
......@@ -968,7 +1107,7 @@ async def oai_chat_completions(user_prompt,
if async_client:
origin = 'llama.cpp'
headers = {'Authorization': f'Bearer {user_api_key}', 'Origin': origin}
async with aiohttp.ClientSession() as session:
async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
async with session.post(f'{base_url}{base_path}',
json=payload,
headers=headers) as response:
......@@ -987,6 +1126,8 @@ async def oai_chat_completions(user_prompt,
event_data = line.split(': ', 1)
assert event_data[0] == 'data', f'Bad event code received: ```{event_data}```'
chunk_raw = event_data[1]
if chunk_raw == '[DONE]':
break
chunk = json.loads(chunk_raw)
assert len(chunk['choices']) == 1, f"no choices provided, line ```{line}```"
......@@ -1054,15 +1195,17 @@ async def oai_chat_completions(user_prompt,
return completion_response
async def request_embedding(content, seed, base_url=None) -> list[list[float]]:
async with aiohttp.ClientSession() as session:
async def request_embedding(content, seed, base_url=None) -> list[list[float]] | int:
async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
async with session.post(f'{base_url}/embedding',
json={
"content": content,
}) as response:
assert response.status == 200
response_json = await response.json()
return [response_json['embedding']]
if response.status == 200:
response_json = await response.json()
return [response_json['embedding']]
else:
return response.status
async def request_oai_embeddings(input, seed,
......@@ -1075,14 +1218,13 @@ async def request_oai_embeddings(input, seed,
headers=[]
if user_api_key is not None:
headers = {'Authorization': f'Bearer {user_api_key}', 'Origin': origin}
async with aiohttp.ClientSession() as session:
async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
async with session.post(f'{base_url}/v1/embeddings',
json={
"input": input,
"model": model,
},
headers=headers,
timeout=3600) as response:
headers=headers) as response:
assert response.status == 200, f"received status code not expected: {response.status}"
assert response.headers['Access-Control-Allow-Origin'] == origin
assert response.headers['Content-Type'] == "application/json; charset=utf-8"
......@@ -1187,44 +1329,36 @@ async def gather_tasks_results(context):
return n_completions
async def wait_for_health_status(context,
base_url,
expected_http_status_code,
expected_health_status,
timeout=3,
params=None,
slots_idle=None,
slots_processing=None,
expected_slots=None):
async def wait_for_slots_status(context,
base_url,
expected_http_status_code,
timeout=3,
params=None,
slots_idle=None,
slots_processing=None):
if context.debug:
print(f"Starting checking for health for expected_health_status={expected_health_status}")
print(f"Starting checking for health for expected_http_status_code={expected_http_status_code}")
interval = 0.5
counter = 0
if 'GITHUB_ACTIONS' in os.environ:
timeout *= 2
async with aiohttp.ClientSession() as session:
async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
while True:
async with await session.get(f'{base_url}/health', params=params) as health_response:
status_code = health_response.status
health = await health_response.json()
headers = {'Authorization': f'Bearer {context.server_api_key}'}
async with await session.get(f'{base_url}/slots', params=params, headers=headers) as slots_response:
status_code = slots_response.status
slots = await slots_response.json()
if context.debug:
print(f"HEALTH - response for expected health status='{expected_health_status}' on "
f"'{base_url}/health'?{params} is {health}\n")
if (status_code == expected_http_status_code
and health['status'] == expected_health_status
and (slots_idle is None or health['slots_idle'] == slots_idle)
and (slots_processing is None or health['slots_processing'] == slots_processing)):
if expected_slots is not None:
assert_slots_status(health['slots'], expected_slots)
return
if (status_code == expected_http_status_code
and health['status'] == expected_health_status
and (slots_idle is None or health['slots_idle'] == slots_idle)
and (slots_processing is None or health['slots_processing'] == slots_processing)):
if expected_slots is not None:
assert_slots_status(health['slots'], expected_slots)
print(f"slots responses {slots}\n")
if status_code == 503 and status_code == expected_http_status_code:
return
if status_code == 200 and status_code == expected_http_status_code:
n_slots_idle = sum(1 if not slot["is_processing"] else 0 for slot in slots)
n_slots_processing = sum(1 if slot["is_processing"] else 0 for slot in slots)
if ((slots_idle is None or slots_idle == n_slots_idle)
and (slots_processing is None or slots_processing == n_slots_processing)):
return
await asyncio.sleep(interval)
counter += interval
......@@ -1238,7 +1372,7 @@ async def wait_for_health_status(context,
if n_completions > 0:
return
assert False, f'{expected_health_status} timeout exceeded {counter}s>={timeout}'
assert False, f'slots check timeout exceeded {counter}s>={timeout}'
def assert_embeddings(embeddings):
......@@ -1253,7 +1387,7 @@ def assert_embeddings(embeddings):
async def request_slots_status(context, expected_slots):
async with aiohttp.ClientSession() as session:
async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
async with await session.get(f'{context.base_url}/slots') as slots_response:
assert slots_response.status == 200
slots = await slots_response.json()
......@@ -1300,6 +1434,7 @@ def start_server_background(context):
context.server_path = os.environ['LLAMA_SERVER_BIN_PATH']
server_listen_addr = context.server_fqdn
server_args = [
'--slots', # requires to get slot status via /slots endpoint
'--host', server_listen_addr,
'--port', context.server_port,
]
......@@ -1325,6 +1460,8 @@ def start_server_background(context):
server_args.append('--cont-batching')
if context.server_embeddings:
server_args.append('--embedding')
if context.server_reranking:
server_args.append('--reranking')
if context.server_metrics:
server_args.append('--metrics')
if context.model_alias:
......@@ -1347,8 +1484,8 @@ def start_server_background(context):
server_args.append('--verbose')
if context.lora_file:
server_args.extend(['--lora', context.lora_file])
if 'SERVER_LOG_FORMAT_JSON' not in os.environ:
server_args.extend(['--log-format', "text"])
if context.disable_ctx_shift:
server_args.extend(['--no-context-shift'])
args = [str(arg) for arg in [context.server_path, *server_args]]
print(f"bench: starting server with: {' '.join(args)}")
......
......@@ -8,9 +8,12 @@ Feature: Wrong usage of llama.cpp server
Scenario: Infinite loop
Given a server listening on localhost:8080
And a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
And 42 as server seed
And 2048 KV cache size
# Uncomment below to fix the issue
#And 64 server max tokens to predict
Then the server is starting
Then the server is healthy
Given a prompt:
"""
Go to: infinite loop
......
aiohttp~=3.9.3
behave~=1.2.6
huggingface_hub~=0.20.3
huggingface_hub~=0.23.2
numpy~=1.26.4
openai~=1.30.3
prometheus-client~=0.20.0
......
......@@ -226,7 +226,6 @@
top_k: 40, // <= 0 to use vocab size
top_p: 0.95, // 1.0 = disabled
min_p: 0.05, // 0 = disabled
tfs_z: 1.0, // 1.0 = disabled
typical_p: 1.0, // 1.0 = disabled
presence_penalty: 0.0, // 0.0 = disabled
frequency_penalty: 0.0, // 0.0 = disabled
......@@ -788,7 +787,6 @@
<details>
<summary>More options</summary>
<fieldset class="two">
${FloatField({ label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
${FloatField({ label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
${FloatField({ label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
${FloatField({ label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}
......
......@@ -229,7 +229,6 @@
top_k: 40, // <= 0 to use vocab size
top_p: 0.95, // 1.0 = disabled
min_p: 0.05, // 0 = disabled
tfs_z: 1.0, // 1.0 = disabled
typical_p: 1.0, // 1.0 = disabled
presence_penalty: 0.0, // 0.0 = disabled
frequency_penalty: 0.0, // 0.0 = disabled
......@@ -791,7 +790,6 @@
<details>
<summary>More options</summary>
<fieldset class="two">
${FloatField({ label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
${FloatField({ label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
${FloatField({ label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
${FloatField({ label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}
......
#pragma once
#include "llama.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#ifndef NDEBUG
// crash the server in debug mode, otherwise send an http 500 error
#define CPPHTTPLIB_NO_EXCEPTIONS 1
#endif
// increase max payload length to allow use of larger context size
#define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 1048576
#include "httplib.h"
// Change JSON_ASSERT from assert() to GGML_ASSERT:
#define JSON_ASSERT GGML_ASSERT
#include "json.hpp"
#include <random>
#include <sstream>
#include <string>
#include <vector>
#include <sstream>
#include <random>
#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
using json = nlohmann::ordered_json;
#define SLT_INF(slot, fmt, ...) LOG_INF("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
#define SLT_WRN(slot, fmt, ...) LOG_WRN("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
#define SLT_ERR(slot, fmt, ...) LOG_ERR("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
#define SLT_DBG(slot, fmt, ...) LOG_DBG("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
#define SRV_INF(fmt, ...) LOG_INF("srv %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define SRV_WRN(fmt, ...) LOG_WRN("srv %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define SRV_ERR(fmt, ...) LOG_ERR("srv %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define SRV_DBG(fmt, ...) LOG_DBG("srv %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define QUE_INF(fmt, ...) LOG_INF("que %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define QUE_WRN(fmt, ...) LOG_WRN("que %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define QUE_ERR(fmt, ...) LOG_ERR("que %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define QUE_DBG(fmt, ...) LOG_DBG("que %12.*s: " fmt, 12, __func__, __VA_ARGS__)
// https://community.openai.com/t/openai-chat-list-of-error-codes-and-types/357791/11
enum error_type {
ERROR_TYPE_INVALID_REQUEST,
......@@ -27,32 +51,6 @@ enum error_type {
ERROR_TYPE_NOT_SUPPORTED, // custom error
};
extern bool server_verbose;
extern bool server_log_json;
#ifndef SERVER_VERBOSE
#define SERVER_VERBOSE 1
#endif
#if SERVER_VERBOSE != 1
#define LOG_VERBOSE(MSG, ...)
#else
#define LOG_VERBOSE(MSG, ...) \
do \
{ \
if (server_verbose) \
{ \
server_log("VERB", __func__, __LINE__, MSG, __VA_ARGS__); \
} \
} while (0)
#endif
#define LOG_ERROR( MSG, ...) server_log("ERR", __func__, __LINE__, MSG, __VA_ARGS__)
#define LOG_WARNING(MSG, ...) server_log("WARN", __func__, __LINE__, MSG, __VA_ARGS__)
#define LOG_INFO( MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__)
static inline void server_log(const char * level, const char * function, int line, const char * message, const json & extra);
template <typename T>
static T json_value(const json & body, const std::string & key, const T & default_value) {
// Fallback null to default value
......@@ -60,9 +58,7 @@ static T json_value(const json & body, const std::string & key, const T & defaul
try {
return body.at(key);
} catch (NLOHMANN_JSON_NAMESPACE::detail::type_error const &) {
std::stringstream ss;
ss << "Wrong type supplied for parameter '" << key << "'. Expected '" << json(default_value).type_name() << "', using default value.";
LOG_WARNING(ss.str().c_str(), body);
LOG_WRN("Wrong type supplied for parameter '%s'. Expected '%s', using default value\n", key.c_str(), json(default_value).type_name());
return default_value;
}
} else {
......@@ -70,55 +66,241 @@ static T json_value(const json & body, const std::string & key, const T & defaul
}
}
static inline void server_log(const char * level, const char * function, int line, const char * message, const json & extra) {
std::stringstream ss_tid;
ss_tid << std::this_thread::get_id();
json log = json{
{"tid", ss_tid.str()},
{"timestamp", time(nullptr)},
};
//
// tokenizer and input processing utils
//
if (server_log_json) {
log.merge_patch({
{"level", level},
{"function", function},
{"line", line},
{"msg", message},
});
static bool json_is_array_of_numbers(const json & data) {
if (data.is_array()) {
for (const auto & e : data) {
if (!e.is_number_integer()) {
return false;
}
}
return true;
}
return false;
}
if (!extra.empty()) {
log.merge_patch(extra);
// is array having BOTH numbers & strings?
static bool json_is_array_of_mixed_numbers_strings(const json & data) {
bool seen_string = false;
bool seen_number = false;
if (data.is_array()) {
for (const auto & e : data) {
seen_string |= e.is_string();
seen_number |= e.is_number_integer();
if (seen_number && seen_string) {
return true;
}
}
}
return false;
}
printf("%s\n", log.dump(-1, ' ', false, json::error_handler_t::replace).c_str());
} else {
char buf[1024];
snprintf(buf, 1024, "%4s [%24s] %s", level, function, message);
/**
* this handles 2 cases:
* - only string, example: "string"
* - mixed string and tokens, example: [12, 34, "string", 56, 78]
*/
static llama_tokens tokenize_mixed(const llama_context * ctx, const json & json_prompt, bool add_special, bool parse_special) {
// If `add_bos` is true, we only add BOS, when json_prompt is a string,
// or the first element of the json_prompt array is a string.
llama_tokens prompt_tokens;
if (json_prompt.is_array()) {
bool first = true;
for (const auto & p : json_prompt) {
if (p.is_string()) {
auto s = p.template get<std::string>();
llama_tokens p;
if (first) {
p = common_tokenize(ctx, s, add_special, parse_special);
first = false;
} else {
p = common_tokenize(ctx, s, false, parse_special);
}
if (!extra.empty()) {
log.merge_patch(extra);
}
std::stringstream ss;
ss << buf << " |";
for (const auto & el : log.items())
{
const std::string value = el.value().dump(-1, ' ', false, json::error_handler_t::replace);
ss << " " << el.key() << "=" << value;
prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end());
} else {
if (first) {
first = false;
}
prompt_tokens.push_back(p.template get<llama_token>());
}
}
} else {
auto s = json_prompt.template get<std::string>();
prompt_tokens = common_tokenize(ctx, s, add_special, parse_special);
}
const std::string str = ss.str();
printf("%.*s\n", (int)str.size(), str.data());
return prompt_tokens;
}
/**
* break the input "prompt" object into multiple prompt if needed, then tokenize them
* this supports these cases:
* - "prompt": "string"
* - "prompt": [12, 34, 56]
* - "prompt": [12, 34, "string", 56, 78]
* and multiple prompts (multi-tasks):
* - "prompt": ["string1", "string2"]
* - "prompt": ["string1", [12, 34, 56]]
* - "prompt": [[12, 34, "string", 56, 78], [12, 34, 56]]
*/
static std::vector<llama_tokens> tokenize_input_prompts(llama_context * ctx, const json & json_prompt, bool add_special, bool parse_special) {
std::vector<llama_tokens> result;
if (json_prompt.is_string() || json_is_array_of_mixed_numbers_strings(json_prompt)) {
// string or mixed
result.push_back(tokenize_mixed(ctx, json_prompt, add_special, parse_special));
} else if (json_is_array_of_numbers(json_prompt)) {
// array of tokens
result.push_back(json_prompt.get<llama_tokens>());
} else if (json_prompt.is_array()) {
// array of prompts
result.reserve(json_prompt.size());
for (const auto & p : json_prompt) {
if (p.is_string() || json_is_array_of_mixed_numbers_strings(p)) {
result.push_back(tokenize_mixed(ctx, p, add_special, parse_special));
} else if (json_is_array_of_numbers(p)) {
// array of tokens
result.push_back(p.get<llama_tokens>());
} else {
throw std::runtime_error("element of \"prompt\" must be a string, an list of tokens, or a list of mixed strings & tokens");
}
}
} else {
throw std::runtime_error("\"prompt\" must be a string, an list of tokens, a list of mixed strings & tokens, or a list of prompts");
}
fflush(stdout);
return result;
}
//
// chat template utils
// template utils
//
// format rerank task: [BOS]query[EOS][SEP]doc[EOS]
static llama_tokens format_rerank(const struct llama_model * model, const llama_tokens & query, const llama_tokens & doc) {
llama_tokens result;
result.reserve(doc.size() + query.size() + 4);
result.push_back(llama_token_bos(model));
result.insert(result.end(), query.begin(), query.end());
result.push_back(llama_token_eos(model));
result.push_back(llama_token_sep(model));
result.insert(result.end(), doc.begin(), doc.end());
result.push_back(llama_token_eos(model));
return result;
}
// format infill task
static llama_tokens format_infill(
const llama_context * ctx,
const json & input_prefix,
const json & input_suffix,
const json & input_extra,
const int n_batch,
const int n_predict,
const int n_ctx,
const bool spm_infill,
const llama_tokens & tokens_prompt
) {
// TODO: optimize this block by reducing memory allocations and movement
// use FIM repo-level pattern:
// ref: https://arxiv.org/pdf/2409.12186
//
// [FIM_REP]myproject
// [FIM_SEP]filename0
// extra chunk 0
// [FIM_SEP]filename1
// extra chunk 1
// ...
// [FIM_SEP]filename
// [FIM_PRE]prefix[FIM_SUF]suffix[FIM_MID]prompt
//
llama_tokens extra_tokens;
extra_tokens.reserve(n_ctx);
auto model = llama_get_model(ctx);
auto tokens_prefix = tokenize_mixed(ctx, input_prefix, false, false);
auto tokens_suffix = tokenize_mixed(ctx, input_suffix, false, false);
if (llama_token_fim_rep(model) != LLAMA_TOKEN_NULL) {
// TODO: make project name an input
static const auto k_fim_repo = common_tokenize(ctx, "myproject\n", false, false);
extra_tokens.push_back(llama_token_fim_rep(model));
extra_tokens.insert(extra_tokens.end(), k_fim_repo.begin(), k_fim_repo.end());
}
for (const auto & chunk : input_extra) {
// { "text": string, "filename": string }
const std::string text = json_value(chunk, "text", std::string());
const std::string filename = json_value(chunk, "filename", std::string("tmp"));
if (llama_token_fim_sep(model) != LLAMA_TOKEN_NULL) {
const auto k_fim_file = common_tokenize(ctx, filename + "\n", false, false);
extra_tokens.insert(extra_tokens.end(), llama_token_fim_sep(model));
extra_tokens.insert(extra_tokens.end(), k_fim_file.begin(), k_fim_file.end());
} else {
// chunk separator in binary form to avoid confusing the AI
static const char k_chunk_prefix_str[] = {0x0a, 0x0a, 0x2d, 0x2d, 0x2d, 0x20, 0x73, 0x6e, 0x69, 0x70, 0x70, 0x65, 0x74, 0x20, 0x2d, 0x2d, 0x2d, 0x0a, 0x0a, 0x00};
static const auto k_chunk_prefix_tokens = common_tokenize(ctx, k_chunk_prefix_str, false, false);
extra_tokens.insert(extra_tokens.end(), k_chunk_prefix_tokens.begin(), k_chunk_prefix_tokens.end());
}
const auto chunk_tokens = common_tokenize(ctx, text, false, false);
extra_tokens.insert(extra_tokens.end(), chunk_tokens.begin(), chunk_tokens.end());
}
if (llama_token_fim_sep(model) != LLAMA_TOKEN_NULL) {
// TODO: current filename
static const auto k_fim_file = common_tokenize(ctx, "filename\n", false, false);
extra_tokens.insert(extra_tokens.end(), llama_token_fim_sep(model));
extra_tokens.insert(extra_tokens.end(), k_fim_file.begin(), k_fim_file.end());
}
// for now pick FIM context to fit in a batch (ratio prefix:suffix = 3:1, TODO: configurable?)
const int n_prefix_take = std::min<int>(tokens_prefix.size(), 3*(n_batch/4));
const int n_suffix_take = std::min<int>(tokens_suffix.size(), std::max<int>(0, (n_batch/4) - (2 + tokens_prompt.size())));
SRV_DBG("n_prefix_take = %d, n_suffix_take = %d, total = %d\n", n_prefix_take, n_suffix_take, (n_prefix_take + n_suffix_take));
// fill the rest of the context with extra chunks
const int n_extra_take = std::min<int>(std::max<int>(0, n_ctx - (n_batch) - 2*n_predict), extra_tokens.size());
tokens_prefix.erase(tokens_prefix.begin(), tokens_prefix.begin() + tokens_prefix.size() - n_prefix_take);
tokens_suffix.resize(n_suffix_take);
tokens_prefix.insert(tokens_prefix.begin(), llama_token_fim_pre(model));
tokens_prefix.insert(tokens_prefix.end(), tokens_prompt.begin(), tokens_prompt.end());
tokens_suffix.insert(tokens_suffix.begin(), llama_token_fim_suf(model));
auto embd_inp = spm_infill ? tokens_suffix : tokens_prefix;
auto embd_end = spm_infill ? tokens_prefix : tokens_suffix;
if (llama_add_bos_token(model)) {
embd_inp.insert(embd_inp.begin(), llama_token_bos(model));
}
SRV_DBG("extra: n_ctx = %d, n_extra_take = %d, n_extra = %d\n", n_ctx, n_extra_take, (int) extra_tokens.size());
// put the extra context before the FIM prefix
embd_inp.insert(embd_inp.begin(), extra_tokens.end() - n_extra_take, extra_tokens.end());
embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
embd_inp.push_back(llama_token_fim_mid(model));
return embd_inp;
}
// Format given chat. If tmpl is empty, we take the template from model metadata
inline std::string format_chat(const struct llama_model * model, const std::string & tmpl, const std::vector<json> & messages) {
std::vector<llama_chat_msg> chat;
std::vector<common_chat_msg> chat;
for (size_t i = 0; i < messages.size(); ++i) {
const auto & curr_msg = messages[i];
......@@ -145,11 +327,25 @@ inline std::string format_chat(const struct llama_model * model, const std::stri
chat.push_back({role, content});
}
auto formatted_chat = llama_chat_apply_template(model, tmpl, chat, true);
LOG_VERBOSE("formatted_chat", {{"text", formatted_chat.c_str()}});
const auto formatted_chat = common_chat_apply_template(model, tmpl, chat, true);
LOG_DBG("formatted_chat: '%s'\n", formatted_chat.c_str());
return formatted_chat;
}
static std::string llama_get_chat_template(const struct llama_model * model) {
std::string template_key = "tokenizer.chat_template";
// call with NULL buffer to get the total size of the string
int32_t res = llama_model_meta_val_str(model, template_key.c_str(), NULL, 0);
if (res < 0) {
return "";
} else {
std::vector<char> model_template(res, 0);
llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), model_template.size());
return std::string(model_template.data(), model_template.size());
}
}
//
// base64 utils (TODO: move to common in the future)
//
......@@ -235,30 +431,13 @@ static std::string random_string() {
}
static std::string gen_chatcmplid() {
std::stringstream chatcmplid;
chatcmplid << "chatcmpl-" << random_string();
return chatcmplid.str();
return "chatcmpl-" + random_string();
}
//
// other common utils
//
static size_t common_part(const std::vector<llama_token> & a, const std::vector<llama_token> & b) {
size_t i;
for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
return i;
}
static size_t common_part(const std::string & a, const std::string & b) {
size_t i;
for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
return i;
}
static bool ends_with(const std::string & str, const std::string & suffix) {
return str.size() >= suffix.size() && 0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix);
}
......@@ -284,7 +463,7 @@ template <class Iter>
static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
std::string ret;
for (; begin != end; ++begin) {
ret += llama_token_to_piece(ctx, *begin);
ret += common_token_to_piece(ctx, *begin);
}
return ret;
......@@ -292,7 +471,7 @@ static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
// format incomplete utf-8 multibyte character for output
static std::string tokens_to_output_formatted_string(const llama_context * ctx, const llama_token token) {
std::string out = token == -1 ? "" : llama_token_to_piece(ctx, token);
std::string out = token == -1 ? "" : common_token_to_piece(ctx, token);
// if the size is 1 and first bit is 1, meaning it's a partial character
// (size > 1 meaning it's already a known token)
......@@ -343,6 +522,17 @@ static json probs_vector_to_json(const llama_context * ctx, const std::vector<co
return out;
}
static bool server_sent_event(httplib::DataSink & sink, const char * event, const json & data) {
const std::string str =
std::string(event) + ": " +
data.dump(-1, ' ', false, json::error_handler_t::replace) +
"\n\n"; // note: these newlines are important (not sure why though, if you know, add a comment to explain)
LOG_DBG("data stream, to_send: %s", str.c_str());
return sink.write(str.c_str(), str.size());
}
//
// OAI utils
//
......@@ -371,6 +561,9 @@ static json oaicompat_completion_params_parse(
std::string response_type = json_value(response_format, "type", std::string());
if (response_type == "json_object") {
llama_params["json_schema"] = json_value(response_format, "schema", json::object());
} else if (response_type == "json_schema") {
json json_schema = json_value(response_format, "json_schema", json::object());
llama_params["json_schema"] = json_value(json_schema, "schema", json::object());
} else if (!response_type.empty() && response_type != "text") {
throw std::runtime_error("response_format type must be one of \"text\" or \"json_object\", but got: " + response_type);
}
......@@ -384,22 +577,22 @@ static json oaicompat_completion_params_parse(
// Handle "logprobs" field
// TODO: The response format of this option is not yet OAI-compatible, but seems like no one really using it; We may need to fix it in the future
if (body.contains("logprobs")) {
if (json_value(body, "logprobs", false)) {
llama_params["n_probs"] = json_value(body, "top_logprobs", 20);
} else if (body.contains("top_logprobs")) {
} else if (body.contains("top_logprobs") && !body.at("top_logprobs").is_null()) {
throw std::runtime_error("top_logprobs requires logprobs to be set to true");
}
// Params supported by OAI but unsupported by llama.cpp
static const std::vector<std::string> unsupported_params { "tools", "tool_choice" };
for (auto & param : unsupported_params) {
for (const auto & param : unsupported_params) {
if (body.contains(param)) {
throw std::runtime_error("Unsupported param: " + param);
}
}
// Copy remaining properties to llama_params
// This allows user to use llama.cpp-specific params like "mirostat", "tfs_z",... via OAI endpoint.
// This allows user to use llama.cpp-specific params like "mirostat", ... via OAI endpoint.
// See "launch_slot_with_task()" for a complete list of params supported by llama.cpp
for (const auto & item : body.items()) {
// Exception: if "n_predict" is present, we overwrite the value specified earlier by "max_tokens"
......@@ -411,7 +604,7 @@ static json oaicompat_completion_params_parse(
return llama_params;
}
static json format_final_response_oaicompat(const json & request, json result, const std::string & completion_id, bool streaming = false) {
static json format_final_response_oaicompat(const json & request, const json & result, const std::string & completion_id, bool streaming = false, bool verbose = false) {
bool stopped_word = result.count("stopped_word") != 0;
bool stopped_eos = json_value(result, "stopped_eos", false);
int num_tokens_predicted = json_value(result, "tokens_predicted", 0);
......@@ -448,7 +641,8 @@ static json format_final_response_oaicompat(const json & request, json result, c
{"id", completion_id}
};
if (server_verbose) {
// extra fields for debugging purposes
if (verbose) {
res["__verbose"] = result;
}
......@@ -460,7 +654,7 @@ static json format_final_response_oaicompat(const json & request, json result, c
}
// return value is vector as there is one case where we might need to generate two responses
static std::vector<json> format_partial_response_oaicompat(json result, const std::string & completion_id) {
static std::vector<json> format_partial_response_oaicompat(const json & result, const std::string & completion_id) {
if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) {
return std::vector<json>({result});
}
......@@ -562,7 +756,7 @@ static std::vector<json> format_partial_response_oaicompat(json result, const st
static json format_embeddings_response_oaicompat(const json & request, const json & embeddings) {
json data = json::array();
int i = 0;
for (auto & elem : embeddings) {
for (const auto & elem : embeddings) {
data.push_back(json{
{"embedding", json_value(elem, "embedding", json::array())},
{"index", i++},
......@@ -573,7 +767,7 @@ static json format_embeddings_response_oaicompat(const json & request, const jso
json res = json {
{"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
{"object", "list"},
{"usage", json {
{"usage", json { // TODO: fill
{"prompt_tokens", 0},
{"total_tokens", 0}
}},
......@@ -583,7 +777,63 @@ static json format_embeddings_response_oaicompat(const json & request, const jso
return res;
}
static json format_tokenizer_response(const std::vector<llama_token> & tokens) {
static json format_response_rerank(const json & request, const json & ranks) {
json data = json::array();
int i = 0;
for (const auto & rank : ranks) {
data.push_back(json{
{"index", i++},
{"relevance_score", json_value(rank, "score", 0.0)},
});
}
json res = json {
{"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
{"object", "list"},
{"usage", json { // TODO: fill
{"prompt_tokens", 0},
{"total_tokens", 0}
}},
{"results", data}
};
return res;
}
static bool is_valid_utf8(const std::string & str) {
const unsigned char* bytes = reinterpret_cast<const unsigned char*>(str.data());
const unsigned char* end = bytes + str.length();
while (bytes < end) {
if (*bytes <= 0x7F) {
// 1-byte sequence (0xxxxxxx)
bytes++;
} else if ((*bytes & 0xE0) == 0xC0) {
// 2-byte sequence (110xxxxx 10xxxxxx)
if (end - bytes < 2 || (bytes[1] & 0xC0) != 0x80)
return false;
bytes += 2;
} else if ((*bytes & 0xF0) == 0xE0) {
// 3-byte sequence (1110xxxx 10xxxxxx 10xxxxxx)
if (end - bytes < 3 || (bytes[1] & 0xC0) != 0x80 || (bytes[2] & 0xC0) != 0x80)
return false;
bytes += 3;
} else if ((*bytes & 0xF8) == 0xF0) {
// 4-byte sequence (11110xxx 10xxxxxx 10xxxxxx 10xxxxxx)
if (end - bytes < 4 || (bytes[1] & 0xC0) != 0x80 ||
(bytes[2] & 0xC0) != 0x80 || (bytes[3] & 0xC0) != 0x80)
return false;
bytes += 4;
} else {
// Invalid UTF-8 lead byte
return false;
}
}
return true;
}
static json format_tokenizer_response(const json & tokens) {
return json {
{"tokens", tokens}
};
......
set(TARGET llama-simple-chat)
add_executable(${TARGET} simple-chat.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
# llama.cpp/example/simple-chat
The purpose of this example is to demonstrate a minimal usage of llama.cpp to create a simple chat program using the chat template from the GGUF file.
```bash
./llama-simple-chat -m Meta-Llama-3.1-8B-Instruct.gguf -c 2048
...
#include "llama.h"
#include <cstdio>
#include <cstring>
#include <iostream>
#include <string>
#include <vector>
static void print_usage(int, char ** argv) {
printf("\nexample usage:\n");
printf("\n %s -m model.gguf [-c context_size] [-ngl n_gpu_layers]\n", argv[0]);
printf("\n");
}
int main(int argc, char ** argv) {
std::string model_path;
int ngl = 99;
int n_ctx = 2048;
// parse command line arguments
for (int i = 1; i < argc; i++) {
try {
if (strcmp(argv[i], "-m") == 0) {
if (i + 1 < argc) {
model_path = argv[++i];
} else {
print_usage(argc, argv);
return 1;
}
} else if (strcmp(argv[i], "-c") == 0) {
if (i + 1 < argc) {
n_ctx = std::stoi(argv[++i]);
} else {
print_usage(argc, argv);
return 1;
}
} else if (strcmp(argv[i], "-ngl") == 0) {
if (i + 1 < argc) {
ngl = std::stoi(argv[++i]);
} else {
print_usage(argc, argv);
return 1;
}
} else {
print_usage(argc, argv);
return 1;
}
} catch (std::exception & e) {
fprintf(stderr, "error: %s\n", e.what());
print_usage(argc, argv);
return 1;
}
}
if (model_path.empty()) {
print_usage(argc, argv);
return 1;
}
// only print errors
llama_log_set([](enum ggml_log_level level, const char * text, void * /* user_data */) {
if (level >= GGML_LOG_LEVEL_ERROR) {
fprintf(stderr, "%s", text);
}
}, nullptr);
// initialize the model
llama_model_params model_params = llama_model_default_params();
model_params.n_gpu_layers = ngl;
llama_model * model = llama_load_model_from_file(model_path.c_str(), model_params);
if (!model) {
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
return 1;
}
// initialize the context
llama_context_params ctx_params = llama_context_default_params();
ctx_params.n_ctx = n_ctx;
ctx_params.n_batch = n_ctx;
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
if (!ctx) {
fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
return 1;
}
// initialize the sampler
llama_sampler * smpl = llama_sampler_chain_init(llama_sampler_chain_default_params());
llama_sampler_chain_add(smpl, llama_sampler_init_min_p(0.05f, 1));
llama_sampler_chain_add(smpl, llama_sampler_init_temp(0.8f));
llama_sampler_chain_add(smpl, llama_sampler_init_dist(LLAMA_DEFAULT_SEED));
// helper function to evaluate a prompt and generate a response
auto generate = [&](const std::string & prompt) {
std::string response;
// tokenize the prompt
const int n_prompt_tokens = -llama_tokenize(model, prompt.c_str(), prompt.size(), NULL, 0, true, true);
std::vector<llama_token> prompt_tokens(n_prompt_tokens);
if (llama_tokenize(model, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), llama_get_kv_cache_used_cells(ctx) == 0, true) < 0) {
GGML_ABORT("failed to tokenize the prompt\n");
}
// prepare a batch for the prompt
llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
llama_token new_token_id;
while (true) {
// check if we have enough space in the context to evaluate this batch
int n_ctx = llama_n_ctx(ctx);
int n_ctx_used = llama_get_kv_cache_used_cells(ctx);
if (n_ctx_used + batch.n_tokens > n_ctx) {
printf("\033[0m\n");
fprintf(stderr, "context size exceeded\n");
exit(0);
}
if (llama_decode(ctx, batch)) {
GGML_ABORT("failed to decode\n");
}
// sample the next token
new_token_id = llama_sampler_sample(smpl, ctx, -1);
// is it an end of generation?
if (llama_token_is_eog(model, new_token_id)) {
break;
}
// convert the token to a string, print it and add it to the response
char buf[256];
int n = llama_token_to_piece(model, new_token_id, buf, sizeof(buf), 0, true);
if (n < 0) {
GGML_ABORT("failed to convert token to piece\n");
}
std::string piece(buf, n);
printf("%s", piece.c_str());
fflush(stdout);
response += piece;
// prepare the next batch with the sampled token
batch = llama_batch_get_one(&new_token_id, 1);
}
return response;
};
std::vector<llama_chat_message> messages;
std::vector<char> formatted(llama_n_ctx(ctx));
int prev_len = 0;
while (true) {
// get user input
printf("\033[32m> \033[0m");
std::string user;
std::getline(std::cin, user);
if (user.empty()) {
break;
}
// add the user input to the message list and format it
messages.push_back({"user", strdup(user.c_str())});
int new_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), true, formatted.data(), formatted.size());
if (new_len > (int)formatted.size()) {
formatted.resize(new_len);
new_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), true, formatted.data(), formatted.size());
}
if (new_len < 0) {
fprintf(stderr, "failed to apply the chat template\n");
return 1;
}
// remove previous messages to obtain the prompt to generate the response
std::string prompt(formatted.begin() + prev_len, formatted.begin() + new_len);
// generate a response
printf("\033[33m");
std::string response = generate(prompt);
printf("\n\033[0m");
// add the response to the messages
messages.push_back({"assistant", strdup(response.c_str())});
prev_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), false, nullptr, 0);
if (prev_len < 0) {
fprintf(stderr, "failed to apply the chat template\n");
return 1;
}
}
// free resources
for (auto & msg : messages) {
free(const_cast<char *>(msg.content));
}
llama_sampler_free(smpl);
llama_free(ctx);
llama_free_model(model);
return 0;
}
set(TARGET llama-simple)
add_executable(${TARGET} simple.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
#include "common.h"
#include "llama.h"
#include <cmath>
#include <cstdio>
#include <cstring>
#include <string>
#include <vector>
static void print_usage(int argc, char ** argv, const gpt_params & params) {
gpt_params_print_usage(argc, argv, params);
LOG_TEE("\nexample usage:\n");
LOG_TEE("\n %s -m model.gguf -p \"Hello my name is\" -n 32\n", argv[0]);
LOG_TEE("\n");
static void print_usage(int, char ** argv) {
printf("\nexample usage:\n");
printf("\n %s -m model.gguf [-n n_predict] [-ngl n_gpu_layers] [prompt]\n", argv[0]);
printf("\n");
}
int main(int argc, char ** argv) {
gpt_params params;
params.prompt = "Hello my name is";
params.n_predict = 32;
if (!gpt_params_parse(argc, argv, params)) {
print_usage(argc, argv, params);
return 1;
// path to the model gguf file
std::string model_path;
// prompt to generate text from
std::string prompt = "Hello my name is";
// number of layers to offload to the GPU
int ngl = 99;
// number of tokens to predict
int n_predict = 32;
// parse command line arguments
{
int i = 1;
for (; i < argc; i++) {
if (strcmp(argv[i], "-m") == 0) {
if (i + 1 < argc) {
model_path = argv[++i];
} else {
print_usage(argc, argv);
return 1;
}
} else if (strcmp(argv[i], "-n") == 0) {
if (i + 1 < argc) {
try {
n_predict = std::stoi(argv[++i]);
} catch (...) {
print_usage(argc, argv);
return 1;
}
} else {
print_usage(argc, argv);
return 1;
}
} else if (strcmp(argv[i], "-ngl") == 0) {
if (i + 1 < argc) {
try {
ngl = std::stoi(argv[++i]);
} catch (...) {
print_usage(argc, argv);
return 1;
}
} else {
print_usage(argc, argv);
return 1;
}
} else {
// prompt starts here
break;
}
}
if (model_path.empty()) {
print_usage(argc, argv);
return 1;
}
if (i < argc) {
prompt = argv[i++];
for (; i < argc; i++) {
prompt += " ";
prompt += argv[i];
}
}
}
// total length of the sequence including the prompt
const int n_predict = params.n_predict;
// init LLM
llama_backend_init();
llama_numa_init(params.numa);
// initialize the model
llama_model_params model_params = llama_model_params_from_gpt_params(params);
llama_model_params model_params = llama_model_default_params();
model_params.n_gpu_layers = ngl;
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
llama_model * model = llama_load_model_from_file(model_path.c_str(), model_params);
if (model == NULL) {
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
return 1;
}
// initialize the context
llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
// tokenize the prompt
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
// find the number of tokens in the prompt
const int n_prompt = -llama_tokenize(model, prompt.c_str(), prompt.size(), NULL, 0, true, true);
if (ctx == NULL) {
fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
// allocate space for the tokens and tokenize the prompt
std::vector<llama_token> prompt_tokens(n_prompt);
if (llama_tokenize(model, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), true, true) < 0) {
fprintf(stderr, "%s: error: failed to tokenize the prompt\n", __func__);
return 1;
}
// tokenize the prompt
std::vector<llama_token> tokens_list;
tokens_list = ::llama_tokenize(ctx, params.prompt, true);
// initialize the context
const int n_ctx = llama_n_ctx(ctx);
const int n_kv_req = tokens_list.size() + (n_predict - tokens_list.size());
llama_context_params ctx_params = llama_context_default_params();
// n_ctx is the context size
ctx_params.n_ctx = n_prompt + n_predict - 1;
// n_batch is the maximum number of tokens that can be processed in a single call to llama_decode
ctx_params.n_batch = n_prompt;
// enable performance counters
ctx_params.no_perf = false;
LOG_TEE("\n%s: n_predict = %d, n_ctx = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, n_kv_req);
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
// make sure the KV cache is big enough to hold all the prompt and generated tokens
if (n_kv_req > n_ctx) {
LOG_TEE("%s: error: n_kv_req > n_ctx, the required KV cache size is not big enough\n", __func__);
LOG_TEE("%s: either reduce n_predict or increase n_ctx\n", __func__);
if (ctx == NULL) {
fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
return 1;
}
// print the prompt token-by-token
fprintf(stderr, "\n");
for (auto id : tokens_list) {
fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
}
// initialize the sampler
fflush(stderr);
auto sparams = llama_sampler_chain_default_params();
sparams.no_perf = false;
llama_sampler * smpl = llama_sampler_chain_init(sparams);
// create a llama_batch with size 512
// we use this object to submit token data for decoding
llama_sampler_chain_add(smpl, llama_sampler_init_greedy());
llama_batch batch = llama_batch_init(512, 0, 1);
// print the prompt token-by-token
// evaluate the initial prompt
for (size_t i = 0; i < tokens_list.size(); i++) {
llama_batch_add(batch, tokens_list[i], i, { 0 }, false);
for (auto id : prompt_tokens) {
char buf[128];
int n = llama_token_to_piece(model, id, buf, sizeof(buf), 0, true);
if (n < 0) {
fprintf(stderr, "%s: error: failed to convert token to piece\n", __func__);
return 1;
}
std::string s(buf, n);
printf("%s", s.c_str());
}
// llama_decode will output logits only for the last token of the prompt
batch.logits[batch.n_tokens - 1] = true;
// prepare a batch for the prompt
if (llama_decode(ctx, batch) != 0) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
return 1;
}
llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
// main loop
int n_cur = batch.n_tokens;
const auto t_main_start = ggml_time_us();
int n_decode = 0;
llama_token new_token_id;
const auto t_main_start = ggml_time_us();
for (int n_pos = 0; n_pos + batch.n_tokens < n_prompt + n_predict; ) {
// evaluate the current batch with the transformer model
if (llama_decode(ctx, batch)) {
fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
return 1;
}
n_pos += batch.n_tokens;
while (n_cur <= n_predict) {
// sample the next token
{
auto n_vocab = llama_n_vocab(model);
auto * logits = llama_get_logits_ith(ctx, batch.n_tokens - 1);
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
}
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
// sample the most likely token
const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
new_token_id = llama_sampler_sample(smpl, ctx, -1);
// is it an end of generation?
if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
LOG_TEE("\n");
if (llama_token_is_eog(model, new_token_id)) {
break;
}
LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
char buf[128];
int n = llama_token_to_piece(model, new_token_id, buf, sizeof(buf), 0, true);
if (n < 0) {
fprintf(stderr, "%s: error: failed to convert token to piece\n", __func__);
return 1;
}
std::string s(buf, n);
printf("%s", s.c_str());
fflush(stdout);
// prepare the next batch
llama_batch_clear(batch);
// push this new token for next evaluation
llama_batch_add(batch, new_token_id, n_cur, { 0 }, true);
// prepare the next batch with the sampled token
batch = llama_batch_get_one(&new_token_id, 1);
n_decode += 1;
}
n_cur += 1;
// evaluate the current batch with the transformer model
if (llama_decode(ctx, batch)) {
fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
return 1;
}
}
LOG_TEE("\n");
printf("\n");
const auto t_main_end = ggml_time_us();
LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
fprintf(stderr, "%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
__func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
llama_print_timings(ctx);
fprintf(stderr, "\n");
llama_perf_sampler_print(smpl);
llama_perf_context_print(ctx);
fprintf(stderr, "\n");
llama_batch_free(batch);
llama_sampler_free(smpl);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
return 0;
}
set(TARGET llama-speculative-simple)
add_executable(${TARGET} speculative-simple.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
# llama.cpp/examples/speculative-simple
Demonstration of basic greedy speculative decoding
```bash
./bin/llama-speculative-simple \
-m ../models/qwen2.5-32b-coder-instruct/ggml-model-q8_0.gguf \
-md ../models/qwen2.5-1.5b-coder-instruct/ggml-model-q4_0.gguf \
-f test.txt -c 0 -ngl 99 --color \
--sampling-seq k --top-k 1 -fa --temp 0.0 \
-ngld 99 --draft-max 16 --draft-min 5 --draft-p-min 0.9
```
#include "arg.h"
#include "common.h"
#include "sampling.h"
#include "speculative.h"
#include "log.h"
#include "llama.h"
#include <cstdio>
#include <cstring>
#include <string>
#include <vector>
int main(int argc, char ** argv) {
common_params params;
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) {
return 1;
}
if (params.n_predict < -1) {
LOG_ERR("%s: --n-predict must be >= -1\n", __func__);
return 1;
}
common_init();
if (params.speculative.model.empty()) {
LOG_ERR("%s: --model-draft is required\n", __func__);
return 1;
}
// init llama.cpp
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model_tgt = NULL;
llama_model * model_dft = NULL;
llama_context * ctx_tgt = NULL;
llama_context * ctx_dft = NULL;
// load the target model
common_init_result llama_init_tgt = common_init_from_params(params);
model_tgt = llama_init_tgt.model;
ctx_tgt = llama_init_tgt.context;
// load the draft model
params.model = params.speculative.model;
params.n_ctx = params.speculative.n_ctx;
params.n_batch = params.speculative.n_ctx > 0 ? params.speculative.n_ctx : params.n_batch;
params.n_gpu_layers = params.speculative.n_gpu_layers;
if (params.speculative.cpuparams.n_threads > 0) {
params.cpuparams.n_threads = params.speculative.cpuparams.n_threads;
}
params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
common_init_result llama_init_dft = common_init_from_params(params);
model_dft = llama_init_dft.model;
ctx_dft = llama_init_dft.context;
if (!common_speculative_are_compatible(ctx_tgt, ctx_dft)) {
return 1;
}
// Tokenize the prompt
std::vector<llama_token> inp;
inp = common_tokenize(ctx_tgt, params.prompt, true, true);
if (llama_n_ctx(ctx_tgt) < (int) inp.size()) {
LOG_ERR("%s: the prompt exceeds the context size (%d tokens, ctx %d)\n", __func__, (int) inp.size(), llama_n_ctx(ctx_tgt));
return 1;
}
if (llama_n_batch(ctx_tgt) < (int) inp.size()) {
LOG_ERR("%s: the prompt exceeds the batch size (%d tokens, batch %d)\n", __func__, (int) inp.size(), llama_n_batch(ctx_tgt));
return 1;
}
LOG("\n\n");
for (auto id : inp) {
LOG("%s", common_token_to_piece(ctx_tgt, id).c_str());
}
// how many tokens to draft each time
int n_draft = params.speculative.n_max;
int n_draft_min = params.speculative.n_min;
float p_min = params.speculative.p_min;
int n_predict = 0;
int n_drafted = 0;
int n_accept = 0;
// used to determine end of generation
bool has_eos = false;
// ================================================
// everything until here is standard initialization
// the relevant stuff for speculative decoding starts here
const auto t_enc_start = ggml_time_us();
// target model sampling context
struct common_sampler * smpl = common_sampler_init(model_tgt, params.sampling);
// eval the prompt
llama_decode(ctx_tgt, llama_batch_get_one(inp.data(), inp.size() - 1));
// note: keep the last token separate!
llama_token id_last = inp.back();
// all tokens currently in the target context
auto prompt_tgt = std::vector<llama_token>(inp.begin(), inp.end() - 1);
int n_past = inp.size() - 1;
// init the speculator
struct common_speculative_params params_spec;
params_spec.n_draft = n_draft;
params_spec.n_reuse = llama_n_ctx(ctx_dft) - n_draft;
params_spec.p_min = p_min;
struct common_speculative * spec = common_speculative_init(ctx_dft);
llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, 1);
const auto t_enc_end = ggml_time_us();
const auto t_dec_start = ggml_time_us();
while (true) {
// optionally, generate draft tokens that can be appended to the target batch
//
// this is the most important part of the speculation. the more probable tokens that are provided here
// the better the performance will be. in theory, this computation can be performed asynchronously and even
// offloaded to a remote device. it doesn't even have to be based on an LLM. instead, it can provide tokens
// from a cache or lookup tables.
//
llama_tokens draft = common_speculative_gen_draft(spec, params_spec, prompt_tgt, id_last);
//LOG_DBG("draft: %s\n", string_from(ctx_dft, draft).c_str());
// always have a token to evaluate from before - id_last
common_batch_clear(batch_tgt);
common_batch_add (batch_tgt, id_last, n_past++, { 0 }, true);
// evaluate the target model on [id_last, draft0, draft1, ..., draftN-1]
{
// do not waste time on small drafts
if (draft.size() < n_draft_min) {
draft.clear();
}
for (size_t i = 0; i < draft.size(); ++i) {
common_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true);
}
//LOG_DBG("target batch: %s\n", string_from(ctx_tgt, batch_tgt).c_str());
llama_decode(ctx_tgt, batch_tgt);
}
// sample from the full target batch and return the accepted tokens based on the target sampler
//
// for each token to be accepted, the sampler would have to sample that same token
// in such cases, instead of decoding the sampled token as we normally do, we simply continue with the
// available logits from the batch and sample the next token until we run out of logits or the sampler
// disagrees with the draft
//
const auto ids = common_sampler_sample_and_accept_n(smpl, ctx_tgt, draft);
//LOG_DBG("ids: %s\n", string_from(ctx_tgt, ids).c_str());
GGML_ASSERT(ids.size() > 0); // there will always be at least one accepted token
n_past += ids.size() - 1;
n_drafted += batch_tgt.n_tokens - 1;
n_accept += ids.size() - 1;
// process the accepted tokens and update contexts
//
// this is the standard token post-processing that we normally do
// in this case, we do it for a group of accepted tokens at once
//
{
llama_token id;
std::string token_str;
for (size_t i = 0; i < ids.size(); ++i) {
id = ids[i];
++n_predict;
if (llama_token_is_eog(model_tgt, id)) {
has_eos = true;
break;
}
token_str = common_token_to_piece(ctx_tgt, id);
if (params.use_color && i + 1 < ids.size()) {
LOG("\u001b[%dm%s\u001b[37m", (36 - 0 % 6), token_str.c_str());
} else {
LOG("%s", token_str.c_str());
}
}
if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
break;
}
LOG_DBG("accepted %d/%d draft tokens, the last target token is: (%d, '%s')\n", (int) ids.size() - 1, (int) draft.size(), id, token_str.c_str());
{
LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past);
llama_kv_cache_seq_rm(ctx_tgt, 0, n_past, -1);
}
prompt_tgt.push_back(id_last);
prompt_tgt.insert(prompt_tgt.end(), ids.begin(), ids.end() - 1);
// remember the last accepted token for the next iteration
id_last = id;
}
}
auto t_dec_end = ggml_time_us();
const int n_input = inp.size();
LOG("\n\n");
LOG_INF("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
LOG_INF("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
LOG_INF("\n");
LOG_INF("n_draft = %d\n", n_draft);
LOG_INF("n_predict = %d\n", n_predict);
LOG_INF("n_drafted = %d\n", n_drafted);
LOG_INF("n_accept = %d\n", n_accept);
LOG_INF("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
LOG_INF("\n");
LOG_INF("draft:\n\n");
llama_perf_context_print(ctx_dft);
LOG_INF("\n");
LOG_INF("target:\n\n");
common_perf_print(ctx_tgt, smpl);
common_sampler_free(smpl);
common_speculative_free(spec);
llama_free(ctx_tgt);
llama_free_model(model_tgt);
llama_free(ctx_dft);
llama_free_model(model_dft);
llama_backend_free();
LOG("\n\n");
return 0;
}
#include "arg.h"
#include "common.h"
#include "sampling.h"
#include "log.h"
#include "llama.h"
#include <cmath>
#include <algorithm>
#include <cstdio>
#include <cstring>
#include <random>
#include <set>
#include <string>
#include <vector>
#include <set>
#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE 100
#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE 128
#define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
struct seq_draft {
......@@ -21,19 +26,28 @@ struct seq_draft {
std::vector<llama_token> tokens;
std::vector<std::vector<llama_token_data>> dists;
struct llama_sampling_context * ctx_sampling;
struct common_sampler * smpl = nullptr;
};
int main(int argc, char ** argv) {
gpt_params params;
common_params params;
// needed to get candidate probs even for temp <= 0.0
params.sampling.n_probs = 128;
if (!gpt_params_parse(argc, argv, params)) {
gpt_params_print_usage(argc, argv, params);
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) {
return 1;
}
if (params.model_draft.empty()) {
fprintf(stderr, "%s: error: --model-draft is required\n", __func__);
if (params.n_predict < -1) {
LOG_ERR("%s: --n-predict must be >= -1\n", __func__);
return 1;
}
common_init();
if (params.speculative.model.empty()) {
LOG_ERR("%s: --model-draft is required\n", __func__);
return 1;
}
......@@ -41,20 +55,11 @@ int main(int argc, char ** argv) {
const int n_seq_dft = params.n_parallel;
// probability threshold for splitting a draft branch (only for n_seq_dft > 1)
const float p_split = params.p_split;
const float p_draft_split = params.speculative.p_split;
if (params.seed == LLAMA_DEFAULT_SEED) {
params.seed = time(NULL);
}
std::default_random_engine rng(params.seed);
std::default_random_engine rng(params.sampling.seed == LLAMA_DEFAULT_SEED ? std::random_device()() : params.sampling.seed);
std::uniform_real_distribution<> u_dist;
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("speculative", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
#endif // LOG_DISABLE_LOGS
// init llama.cpp
llama_backend_init();
llama_numa_init(params.numa);
......@@ -66,30 +71,31 @@ int main(int argc, char ** argv) {
llama_context * ctx_dft = NULL;
// load the target model
llama_init_result llama_init_tgt = llama_init_from_gpt_params(params);
common_init_result llama_init_tgt = common_init_from_params(params);
model_tgt = llama_init_tgt.model;
ctx_tgt = llama_init_tgt.context;
// load the draft model
params.model = params.model_draft;
params.n_gpu_layers = params.n_gpu_layers_draft;
if (params.n_threads_draft > 0) {
params.n_threads = params.n_threads_draft;
params.model = params.speculative.model;
params.n_gpu_layers = params.speculative.n_gpu_layers;
if (params.speculative.cpuparams.n_threads > 0) {
params.cpuparams.n_threads = params.speculative.cpuparams.n_threads;
}
params.n_threads_batch = params.n_threads_batch_draft;
llama_init_result llama_init_dft = llama_init_from_gpt_params(params);
params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
common_init_result llama_init_dft = common_init_from_params(params);
model_dft = llama_init_dft.model;
ctx_dft = llama_init_dft.context;
const bool vocab_type_tgt = llama_vocab_type(model_tgt);
LOG("vocab_type tgt: %d\n", vocab_type_tgt);
LOG_DBG("vocab_type tgt: %d\n", vocab_type_tgt);
const bool vocab_type_dft = llama_vocab_type(model_dft);
LOG("vocab_type dft: %d\n", vocab_type_dft);
LOG_DBG("vocab_type dft: %d\n", vocab_type_dft);
if (vocab_type_tgt != vocab_type_dft) {
fprintf(stderr, "%s: error: draft model vocab type must match target model to use speculation but ", __func__);
fprintf(stderr, "vocab_type_dft = %d while vocab_type_tgt = %d\n", vocab_type_dft, vocab_type_tgt);
LOG_ERR("%s: draft model vocab type must match target model to use speculation but ", __func__);
LOG_ERR("vocab_type_dft = %d while vocab_type_tgt = %d\n", vocab_type_dft, vocab_type_tgt);
return 1;
}
......@@ -99,7 +105,7 @@ int main(int argc, char ** argv) {
llama_token_bos(model_tgt) != llama_token_bos(model_dft) ||
llama_token_eos(model_tgt) != llama_token_eos(model_dft)
) {
fprintf(stderr, "%s: error: draft model special tokens must match target model to use speculation\n", __func__);
LOG_ERR("%s: draft model special tokens must match target model to use speculation\n", __func__);
return 1;
}
......@@ -111,8 +117,8 @@ int main(int argc, char ** argv) {
: n_vocab_dft - n_vocab_tgt;
if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
fprintf(stderr, "%s: error: draft model vocab must closely match target model to use speculation but ", __func__);
fprintf(stderr, "target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
LOG_ERR("%s: draft model vocab must closely match target model to use speculation but ", __func__);
LOG_ERR("target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
return 1;
}
......@@ -121,10 +127,10 @@ int main(int argc, char ** argv) {
const char * token_text_tgt = llama_token_get_text(model_tgt, i);
const char * token_text_dft = llama_token_get_text(model_dft, i);
if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
fprintf(stderr, "%s: error: draft model vocab must match target model to use speculation but ", __func__);
fprintf(stderr, "token %d content differs - target '%s', draft '%s'\n", i,
llama_token_to_piece(ctx_tgt, i).c_str(),
llama_token_to_piece(ctx_dft, i).c_str());
LOG_ERR("%s: draft model vocab must match target model to use speculation but ", __func__);
LOG_ERR("token %d content differs - target '%s', draft '%s'\n", i,
common_token_to_piece(ctx_tgt, i).c_str(),
common_token_to_piece(ctx_dft, i).c_str());
return 1;
}
}
......@@ -133,32 +139,30 @@ int main(int argc, char ** argv) {
// Tokenize the prompt
std::vector<llama_token> inp;
inp = ::llama_tokenize(ctx_tgt, params.prompt, true, true);
inp = common_tokenize(ctx_tgt, params.prompt, true, true);
const int max_context_size = llama_n_ctx(ctx_tgt);
const int max_tokens_list_size = max_context_size - 4;
if ((int) inp.size() > max_tokens_list_size) {
fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
LOG_ERR("%s: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
return 1;
}
fprintf(stderr, "\n\n");
LOG("\n\n");
for (auto id : inp) {
fprintf(stderr, "%s", llama_token_to_piece(ctx_tgt, id).c_str());
LOG("%s", common_token_to_piece(ctx_tgt, id).c_str());
}
fflush(stderr);
const int n_input = inp.size();
const auto t_enc_start = ggml_time_us();
// eval the prompt with both models
llama_decode(ctx_tgt, llama_batch_get_one( inp.data(), n_input - 1, 0, 0));
llama_decode(ctx_tgt, llama_batch_get_one(&inp.back(), 1, n_input - 1, 0));
llama_decode(ctx_dft, llama_batch_get_one( inp.data(), n_input, 0, 0));
llama_decode(ctx_tgt, llama_batch_get_one( inp.data(), n_input - 1));
llama_decode(ctx_tgt, llama_batch_get_one(&inp.back(), 1));
llama_decode(ctx_dft, llama_batch_get_one( inp.data(), n_input));
const auto t_enc_end = ggml_time_us();
......@@ -166,7 +170,7 @@ int main(int argc, char ** argv) {
//GGML_ASSERT(n_vocab == llama_n_vocab(model_dft));
// how many tokens to draft each time
int n_draft = params.n_draft;
int n_draft = params.speculative.n_max;
int n_predict = 0;
int n_drafted = 0;
......@@ -178,23 +182,19 @@ int main(int argc, char ** argv) {
// used to determine end of generation
bool has_eos = false;
// target model sampling context
struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams);
// target model sampling context (reuse the llama_context's sampling instance)
struct common_sampler * smpl = common_sampler_init(model_tgt, params.sampling);
// draft sequence data
std::vector<seq_draft> drafts(n_seq_dft);
params.sparams.grammar.clear(); // the draft samplers will copy the target sampler's grammar
if (params.sparams.temp == 0) {
params.sparams.temp = -1.0f; // force greedy sampling with probs for the draft model
}
for (int s = 0; s < n_seq_dft; ++s) {
drafts[s].ctx_sampling = llama_sampling_init(params.sparams);
// allocate llama_sampler for each draft sequence
drafts[s].smpl = common_sampler_init(model_dft, params.sampling);
}
llama_batch batch_dft = llama_batch_init(params.n_ctx, 0, 1);
llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, n_seq_dft);
llama_batch batch_dft = llama_batch_init(llama_n_batch(ctx_dft), 0, 1);
llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, n_seq_dft);
const auto t_dec_start = ggml_time_us();
......@@ -214,7 +214,7 @@ int main(int argc, char ** argv) {
active_seqs.insert(s);
const auto & tokens = drafts[s].tokens;
LOG("draft %d: %s\n", s, LOG_TOKENS_TOSTR_PRETTY(ctx_dft, tokens).c_str());
LOG_DBG("draft %d: %s\n", s, string_from(ctx_dft, tokens).c_str());
}
int i_dft = 0;
......@@ -230,14 +230,14 @@ int main(int argc, char ** argv) {
// for stochastic sampling, attempt to match the token with the drafted tokens
{
bool accept = false;
if (params.sparams.temp > 0) {
if (params.sampling.temp > 0) {
// stochastic verification
common_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft], true);
llama_token_data_array dist_tgt = llama_sampling_prepare(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft], true, NULL);
llama_sample_softmax(ctx_tgt, &dist_tgt);
float p_tgt = 0, p_dft = 0;
auto & dist_tgt = *common_sampler_get_candidates(smpl);
// GGML_ASSERT(dist_tgt.size() == dist_dft.size());
float p_tgt = 0.0f;
float p_dft = 0.0f;
while (active_seqs.size() > 0) {
// randomly select a sequence to verify from active sequences
......@@ -256,39 +256,43 @@ int main(int argc, char ** argv) {
}
continue;
}
LOG("verifying sequence #%d at pos #%d from %d active sequence(s)\n", s, i_dft, (int) active_seqs.size());
LOG_DBG("verifying sequence #%d at pos #%d from %d active sequence(s)\n", s, i_dft, (int) active_seqs.size());
float r = u_dist(rng);
llama_token_data_array dist_dft = { drafts[s].dists[i_dft].data() , drafts[s].dists[i_dft].size(), true };
llama_token_data_array dist_dft = { drafts[s].dists[i_dft].data() , drafts[s].dists[i_dft].size(), LLAMA_TOKEN_NULL, true };
//GGML_ASSERT(dist_tgt.size <= dist_dft.size);
// acquire the token probabilities assigned by the draft and target models
for (size_t i = 0; i < dist_tgt.size; i++) {
if (dist_tgt.data[i].id == drafts[s].tokens[i_dft]) {
p_tgt = dist_tgt.data[i].p;
break;
}
}
for (size_t i = 0; i < dist_dft.size; i++) {
if (dist_dft.data[i].id == drafts[s].tokens[i_dft]) {
p_dft = dist_dft.data[i].p;
}
if (p_tgt && p_dft) {
break;
}
}
LOG("r = %f, p_dft = %f, p_tgt = %f\n", r, p_dft, p_tgt);
LOG_DBG("r = %f, p_dft = %f, p_tgt = %f\n", r, p_dft, p_tgt);
if (r <= p_tgt / p_dft) {
s_keep = s;
accept = true;
token_id = drafts[s].tokens[i_dft];
token_str = llama_token_to_piece(ctx_tgt, token_id);
llama_sampling_accept(ctx_sampling, ctx_tgt, token_id, true);
token_str = common_token_to_piece(ctx_tgt, token_id);
common_sampler_accept(smpl, token_id, true);
LOG("draft token %d of sequence %d (%d, '%s') accepted\n", i_dft, s, token_id, token_str.c_str());
LOG_DBG("draft token %d of sequence %d (%d, '%s') accepted\n", i_dft, s, token_id, token_str.c_str());
break;
} else {
LOG("draft token %d of sequence %d (%d, '%s') rejected\n", i_dft, s, drafts[s].tokens[i_dft], llama_token_to_piece(ctx_tgt, drafts[s].tokens[i_dft]).c_str());
LOG_DBG("draft token %d of sequence %d (%d, '%s') rejected\n", i_dft, s, drafts[s].tokens[i_dft], common_token_to_piece(ctx_tgt, drafts[s].tokens[i_dft]).c_str());
drafts[s].active = false;
// calculate residual probability
GGML_ASSERT(dist_tgt.sorted);
GGML_ASSERT(dist_dft.sorted);
float sum_probs = 0.0f;
// sort dist by id
std::sort(dist_tgt.data, dist_tgt.data + dist_tgt.size, [](const llama_token_data &a, const llama_token_data &b) {
......@@ -298,10 +302,18 @@ int main(int argc, char ** argv) {
return a.id < b.id;
});
float sum_probs = 0.0f;
for (size_t i = 0; i < dist_tgt.size; i++) {
dist_tgt.data[i].p = std::max(0.0f, dist_tgt.data[i].p - dist_dft.data[i].p);
if (i < dist_dft.size) {
dist_tgt.data[i].p = std::max(0.0f, dist_tgt.data[i].p - dist_dft.data[i].p);
} else {
dist_tgt.data[i].p = std::max(0.0f, dist_tgt.data[i].p);
}
sum_probs += dist_tgt.data[i].p;
}
for (size_t i = 0; i < dist_tgt.size; i++) {
dist_tgt.data[i].p /= sum_probs;
}
......@@ -330,24 +342,30 @@ int main(int argc, char ** argv) {
if (!accept) {
// all drafted tokens were rejected
// sample from the target model
LOG("all drafted tokens were rejected, sampling from residual distribution\n");
token_id = llama_sample_token(ctx_tgt, &dist_tgt);
llama_sampling_accept(ctx_sampling, ctx_tgt, token_id, true);
token_str = llama_token_to_piece(ctx_tgt, token_id);
}
LOG_DBG("all drafted tokens were rejected, sampling from residual distribution\n");
std::vector<float> probs(dist_tgt.size);
for (size_t i = 0; i < dist_tgt.size; ++i) {
probs[i] = dist_tgt.data[i].p;
}
std::discrete_distribution<> dist(probs.begin(), probs.end());
const int idx = dist(rng);
token_id = dist_tgt.data[idx].id;
common_sampler_accept(smpl, token_id, true);
token_str = common_token_to_piece(ctx_tgt, token_id);
}
} else {
// greedy verification
// sample from the target model
LOG("sampling target: s_keep = %3d, i_dft = %3d, i_batch_tgt = %3d\n", s_keep, i_dft, drafts[s_keep].i_batch_tgt[i_dft]);
token_id = llama_sampling_sample(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft]);
llama_sampling_accept(ctx_sampling, ctx_tgt, token_id, true);
LOG_DBG("sampling target: s_keep = %3d, i_dft = %3d, i_batch_tgt = %3d\n", s_keep, i_dft, drafts[s_keep].i_batch_tgt[i_dft]);
token_id = common_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft]);
//LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, ctx_sampling->prev).c_str());
common_sampler_accept(smpl, token_id, true);
token_str = llama_token_to_piece(ctx_tgt, token_id);
token_str = common_token_to_piece(ctx_tgt, token_id);
for (int s = 0; s < n_seq_dft; ++s) {
if (!drafts[s].active) {
......@@ -355,7 +373,7 @@ int main(int argc, char ** argv) {
}
if (i_dft < (int) drafts[s].tokens.size() && token_id == drafts[s].tokens[i_dft]) {
LOG("the sampled target token matches the %dth drafted token of sequence %d (%d, '%s') - accepted\n", i_dft, s, token_id, token_str.c_str());
LOG_DBG("the sampled target token matches the %dth drafted token of sequence %d (%d, '%s') - accepted\n", i_dft, s, token_id, token_str.c_str());
s_keep = s;
accept = true;
......@@ -377,26 +395,24 @@ int main(int argc, char ** argv) {
++i_dft;
if (params.use_color) {
// Color token according to its origin sequence
printf("\u001b[%dm%s\u001b[37m", (36 - s_keep % 6), token_str.c_str());
LOG("\u001b[%dm%s\u001b[37m", (36 - s_keep % 6), token_str.c_str());
} else {
printf("%s", token_str.c_str());
LOG("%s", token_str.c_str());
}
fflush(stdout);
continue;
} else {
printf("%s", token_str.c_str());
fflush(stdout);
LOG("%s", token_str.c_str());
break;
}
}
}
{
LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", token_id, token_str.c_str());
LOG_DBG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", token_id, token_str.c_str());
// TODO: simplify
{
LOG("keeping sequence %d, n_past_tgt = %d, n_past_dft = %d\n", s_keep, n_past_tgt, n_past_dft);
LOG_DBG("keeping sequence %d, n_past_tgt = %d, n_past_dft = %d\n", s_keep, n_past_tgt, n_past_dft);
llama_kv_cache_seq_keep(ctx_dft, s_keep);
llama_kv_cache_seq_cp (ctx_dft, s_keep, 0, -1, -1);
......@@ -419,21 +435,24 @@ int main(int argc, char ** argv) {
drafts[0].dists.push_back(std::vector<llama_token_data>());
drafts[0].i_batch_tgt.push_back(0);
llama_batch_clear(batch_dft);
llama_batch_add (batch_dft, token_id, n_past_dft, { 0 }, true);
common_batch_clear(batch_dft);
common_batch_add (batch_dft, token_id, n_past_dft, { 0 }, true);
llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
// LOG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
// LOG_DBG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
llama_decode(ctx_dft, batch_dft);
++n_past_dft;
}
if (n_predict > params.n_predict || has_eos) {
if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
break;
}
llama_sampling_cp(ctx_sampling, drafts[0].ctx_sampling);
if (drafts[0].smpl) {
common_sampler_free(drafts[0].smpl);
}
drafts[0].smpl = common_sampler_clone(smpl);
int n_seq_cur = 1;
int n_past_cur = n_past_dft;
......@@ -446,8 +465,8 @@ int main(int argc, char ** argv) {
drafts[0].drafting = true;
drafts[0].i_batch_dft = 0;
llama_batch_clear(batch_tgt);
llama_batch_add (batch_tgt, drafts[0].tokens[0], n_past_tgt, { 0 }, true);
common_batch_clear(batch_tgt);
common_batch_add (batch_tgt, drafts[0].tokens[0], n_past_tgt, { 0 }, true);
// sample n_draft tokens from the draft model using tree-based sampling
for (int i = 0; i < n_draft; ++i) {
......@@ -462,21 +481,21 @@ int main(int argc, char ** argv) {
continue;
}
llama_sampling_sample(drafts[s].ctx_sampling, ctx_dft, NULL, drafts[s].i_batch_dft);
common_sampler_sample(drafts[s].smpl, ctx_dft, drafts[s].i_batch_dft, true);
const auto & cur_p = drafts[s].ctx_sampling->cur;
const auto * cur_p = common_sampler_get_candidates(drafts[s].smpl);
for (int k = 0; k < std::min(n_seq_dft + 3, (int) cur_p.size()); ++k) {
LOG(" - draft candidate %3d for seq %3d, pos %3d: %6d (%8.3f) '%s'\n",
k, s, i, cur_p[k].id, cur_p[k].p, llama_token_to_piece(ctx_dft, cur_p[k].id).c_str());
for (int k = 0; k < std::min(n_seq_dft + 3, (int) cur_p->size); ++k) {
LOG_DBG(" - draft candidate %3d for seq %3d, pos %3d: %6d (%8.3f) '%s'\n",
k, s, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
}
std::vector<int> sa(1, s);
// attempt to split the branch if the probability is high enough
for (int f = 1; f < 8; ++f) {
if (n_seq_cur < n_seq_dft && cur_p[f].p > p_split) {
LOG("splitting seq %3d into %3d\n", s, n_seq_cur);
if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_draft_split) {
LOG_DBG("splitting seq %3d into %3d\n", s, n_seq_cur);
llama_kv_cache_seq_rm(ctx_dft, n_seq_cur, -1, -1);
llama_kv_cache_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
......@@ -502,7 +521,10 @@ int main(int argc, char ** argv) {
drafts[n_seq_cur].i_batch_dft = drafts[s].i_batch_dft;
drafts[n_seq_cur].i_batch_tgt = drafts[s].i_batch_tgt;
llama_sampling_cp(drafts[s].ctx_sampling, drafts[n_seq_cur].ctx_sampling);
if (drafts[n_seq_cur].smpl) {
common_sampler_free(drafts[n_seq_cur].smpl);
}
drafts[n_seq_cur].smpl = common_sampler_clone(drafts[s].smpl);
sa.push_back(n_seq_cur);
......@@ -514,25 +536,25 @@ int main(int argc, char ** argv) {
// add drafted token for each sequence
for (int is = 0; is < (int) sa.size(); ++is) {
const llama_token id = cur_p[is].id;
const llama_token id = cur_p->data[is].id;
const int s = sa[is];
llama_sampling_accept(drafts[s].ctx_sampling, ctx_dft, id, true);
common_sampler_accept(drafts[s].smpl, id, true);
drafts[s].tokens.push_back(id);
// save cur_p.data into drafts[s].dists
drafts[s].dists.push_back(cur_p);
drafts[s].dists.push_back({cur_p->data, cur_p->data + cur_p->size});
// add unique drafted tokens to the target batch
drafts[s].i_batch_tgt.push_back(batch_tgt.n_tokens);
llama_batch_add(batch_tgt, id, n_past_tgt + i + 1, { s }, true);
common_batch_add(batch_tgt, id, n_past_tgt + i + 1, { s }, true);
// add the token to the batch for batched decoding with the draft model
drafts[s].i_batch_dft = batch_dft.n_tokens;
llama_batch_add(batch_dft, id, n_past_cur, { s }, true);
common_batch_add(batch_dft, id, n_past_cur, { s }, true);
if (batch_tgt.n_tokens > n_draft) {
drafts[s].drafting = false;
......@@ -562,7 +584,7 @@ int main(int argc, char ** argv) {
llama_kv_cache_seq_cp(ctx_tgt, 0, s, -1, -1);
}
// LOG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str());
// LOG_DBG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str());
llama_decode(ctx_tgt, batch_tgt);
++n_past_tgt;
}
......@@ -580,27 +602,30 @@ int main(int argc, char ** argv) {
auto t_dec_end = ggml_time_us();
LOG_TEE("\n\n");
LOG("\n\n");
LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
LOG_INF("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
LOG_INF("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
LOG_TEE("\n");
LOG_TEE("n_draft = %d\n", n_draft);
LOG_TEE("n_predict = %d\n", n_predict);
LOG_TEE("n_drafted = %d\n", n_drafted);
LOG_TEE("n_accept = %d\n", n_accept);
LOG_TEE("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
LOG_INF("\n");
LOG_INF("n_draft = %d\n", n_draft);
LOG_INF("n_predict = %d\n", n_predict);
LOG_INF("n_drafted = %d\n", n_drafted);
LOG_INF("n_accept = %d\n", n_accept);
LOG_INF("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
LOG_TEE("\ndraft:\n");
llama_print_timings(ctx_dft);
LOG_INF("\n");
LOG_INF("draft:\n\n");
// TODO: print sampling/grammar timings for all drafts
llama_perf_context_print(ctx_dft);
LOG_TEE("\ntarget:\n");
llama_print_timings(ctx_tgt);
LOG_INF("\n");
LOG_INF("target:\n\n");
common_perf_print(ctx_tgt, smpl);
llama_sampling_free(ctx_sampling);
common_sampler_free(smpl);
for (int s = 0; s < n_seq_dft; ++s) {
llama_sampling_free(drafts[s].ctx_sampling);
common_sampler_free(drafts[s].smpl);
}
llama_batch_free(batch_dft);
......@@ -613,7 +638,7 @@ int main(int argc, char ** argv) {
llama_backend_free();
fprintf(stderr, "\n\n");
LOG("\n\n");
return 0;
}
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment