Skip to content

GitLab

Unverified Commit 5a7fb3ab authored by Asher's avatar Asher Committed by GitHub
Browse files

[Model] Add ToolParser and MoE Config for Hunyuan A13B (#20820) 


Signed-off-by: default avatarAsher Zhang <asherszhang@tencent.com>
parent 11dfdf21
Show whitespace changes
Inline Side-by-side
Showing with 1712 additions and 4 deletions
benchmarks/kernels/benchmark_moe.py
View file @ 5a7fb3ab
...@@ -586,6 +586,11 @@ def main(args: argparse.Namespace): ...@@ -586,6 +586,11 @@ def main(args: argparse.Namespace):
topk = config.num_experts_per_tok topk = config.num_experts_per_tok
intermediate_size = config.moe_intermediate_size intermediate_size = config.moe_intermediate_size
shard_intermediate_size = 2 * intermediate_size // args.tp_size shard_intermediate_size = 2 * intermediate_size // args.tp_size
elif config.architectures[0] in ("HunYuanMoEV1ForCausalLM"):
E = config.num_experts
topk = config.moe_topk[0]
intermediate_size = config.moe_intermediate_size[0]
shard_intermediate_size = 2 * intermediate_size // args.tp_size
else: else:
# Support for llama4 # Support for llama4
config = config.get_text_config() config = config.get_text_config()
......
docs/features/reasoning_outputs.md
View file @ 5a7fb3ab
...@@ -14,6 +14,7 @@ vLLM currently supports the following reasoning models: ...@@ -14,6 +14,7 @@ vLLM currently supports the following reasoning models:
| [QwQ-32B](https://huggingface.co/Qwen/QwQ-32B) | `deepseek_r1` | `guided_json`, `guided_regex` | ✅ | | [QwQ-32B](https://huggingface.co/Qwen/QwQ-32B) | `deepseek_r1` | `guided_json`, `guided_regex` | ✅ |
| [IBM Granite 3.2 language models](https://huggingface.co/collections/ibm-granite/granite-32-language-models-67b3bc8c13508f6d064cff9a) | `granite` | ❌ | ❌ | | [IBM Granite 3.2 language models](https://huggingface.co/collections/ibm-granite/granite-32-language-models-67b3bc8c13508f6d064cff9a) | `granite` | ❌ | ❌ |
| [Qwen3 series](https://huggingface.co/collections/Qwen/qwen3-67dd247413f0e2e4f653967f) | `qwen3` | `guided_json`, `guided_regex` | ✅ | | [Qwen3 series](https://huggingface.co/collections/Qwen/qwen3-67dd247413f0e2e4f653967f) | `qwen3` | `guided_json`, `guided_regex` | ✅ |
| [Hunyuan A13B series](https://huggingface.co/collections/tencent/hunyuan-a13b-685ec38e5b46321e3ea7c4be) | `hunyuan_a13b` | `guided_json`, `guided_regex` | ✅ |
!!! note !!! note
IBM Granite 3.2 reasoning is disabled by default; to enable it, you must also pass `thinking=True` in your `chat_template_kwargs`. IBM Granite 3.2 reasoning is disabled by default; to enable it, you must also pass `thinking=True` in your `chat_template_kwargs`.
......
docs/features/tool_calling.md
View file @ 5a7fb3ab
...@@ -288,6 +288,16 @@ Supported models: ...@@ -288,6 +288,16 @@ Supported models:
Flags: `--tool-call-parser kimi_k2` Flags: `--tool-call-parser kimi_k2`
### Hunyuan Models (`hunyuan_a13b`)
Supported models:
* `tencent/Hunyuan-A13B-Instruct` (chat template already included huggingface model file.)
Flags:
* For non-reasoning: `--tool-call-parser hunyuan_a13b`
* For reasoning: `--tool-call-parser hunyuan_a13b --reasoning-parser hunyuan_a13b --enable_reasoning`
### Models with Pythonic Tool Calls (`pythonic`) ### Models with Pythonic Tool Calls (`pythonic`)
A growing number of models output a python list to represent tool calls instead of using JSON. This has the advantage of inherently supporting parallel tool calls and removing ambiguity around the JSON schema required for tool calls. The `pythonic` tool parser can support such models. A growing number of models output a python list to represent tool calls instead of using JSON. This has the advantage of inherently supporting parallel tool calls and removing ambiguity around the JSON schema required for tool calls. The `pythonic` tool parser can support such models.
......
examples/tool_chat_template_hunyuan_a13b.jinja 0 → 100644
View file @ 5a7fb3ab
{% set loop_messages = messages %}
{% if tools %}
{% set weekday_map = {'Monday': '星期一', 'Tuesday': '星期二', 'Wednesday': '星期三', 'Thursday': '星期四', 'Friday': '星期五', 'Saturday': '星期六', 'Sunday': '星期日'} %}
{% set weekday_cn = weekday_map[strftime_now('%A')] %}
{% set datetime_str = strftime_now('%Y-%m-%d %H:%M:%S') %}
{% set datetime_str = datetime_str + ' ' + weekday_cn %}
{% for message in loop_messages %}
{% if 'content' in message %}
{% set content = message['content'] %}
{% else %}
{% set content = '' %}
{% endif %}
{% if loop.index0 == 0 %}
{% set content_tmp = '你是一位函数组合专家。你会得到一个问题和一组可能的函数。根据问题，你需要进行一个或多个函数/工具调用以实现目的。
如果没有一个函数可以使用，请直接使用自然语言回复用户，以助手：开头。
如果给定的问题缺少函数所需的参数，请使用自然语言进行提问，向用户询问必要信息，以助手：开头。
如果调用结果已经足够回答用户问题，请对历史结果进行总结，使用自然语言回复用户，以助手：开头。
你应该只在工具调用部分返回函数调用。如果你决定调用任何函数，你必须将其格式化为<tool_calls>[{"name": "func_name1", "arguments": {"argument1": "value1", "argument2": "value2"}},...]</tool_calls>。你不应该在回复中包含任何其他文本。以下是你可以调用的函数列表，格式为JSON。
' %}
{% set content_tmp = content_tmp + '
' + tools | tojson + '
' %}
{% if message['role'] == 'system' %}
{% set content_tmp = content_tmp + '
额外要求：
' + content + '
如果你决定返回函数调用，请将其格式化为<tool_calls>[{"name": "func_name1", "arguments": {"argument1": "value1", "argument2": "value2"}},...]</tool_calls>，不得包含其他文本。如果额外要求里有格式要求，请忽略，以此处为准。
否则，请参考开头说的三种情况，以助手：开头进行回复。
如果额外要求里有时间信息，就以额外要求里的时间为准，否则，参考当前时间：' + datetime_str %}
{% set content = '<|startoftext|>' + content_tmp + '<|extra_4|>' %}
{% elif message['role'] == 'user' %}
{% set content_tmp = content_tmp + '
如果你决定返回函数调用，请将其格式化为<tool_calls>[{"name": "func_name1", "arguments": {"argument1": "value1", "argument2": "value2"}},...]</tool_calls>，不得包含其他文本。
否则，请参考开头说的三种情况，以助手：开头进行回复。
当前时间：' + datetime_str %}
{% set content_tmp = '<|startoftext|>' + content_tmp + '<|extra_4|>'%}
{% set content = content_tmp + '用户：' + content + '<|extra_0|>' %}
{% endif %}
{% else %}
{% if message['role'] == 'user' %}
{% set content = '用户：' + content + '<|extra_0|>' %}
{% elif message['role'] == 'assistant' %}
{% if 'tool_calls' in message %}
{% set tool_calls = message['tool_calls'] %}
{% set ns = namespace(tool_calls="[") %}
{% for tool_call in tool_calls %}
{% set function = tool_call['function'] %}
{% set name = function['name'] %}
{% set ns.tool_calls = ns.tool_calls + '{"name": "' + name + '", '%}
{% set arguments = function['arguments'] %}
{% if arguments is not string %}
{% set arguments = arguments | tojson %}
{% endif %}
{% set ns.tool_calls = ns.tool_calls + '"arguments": ' + arguments + '}' %}
{% if not loop.last %}
{% set ns.tool_calls = ns.tool_calls + ', '%}
{% endif %}
{% endfor %}
{% set ns.tool_calls = ns.tool_calls + ']' %}
{% set content = content + '<tool_calls>' + ns.tool_calls + '</tool_calls>' %}
{% else %}
{% set content = '助手：' + content %}
{% endif %}
{% set content = content + '<|eos|>' %}
{% elif message['role'] == 'tool' %}
{% if content is not string %}
{set content = content | tojson }
{% endif %}
{% set content = '<tool_response>' + content + '</tool_response>' %}
{% set content = content + '<|extra_0|>' %}
{% endif %}
{% endif %}
{{- content -}}
{% endfor %}
{% else %}
{% set context = {'has_head': true} %}
{% for message in loop_messages %}
{% if 'content' in message %}
{% set content = message['content'] %}
{% else %}
{% set content = '' %}
{% endif %}
{% if loop.index0 == 0 %}
{% if content == '' %}
{% set _ = context.update({'has_head': false}) %}
{% elif message['role'] == 'system' %}
{% set content = '<|startoftext|>' + content + '<|extra_4|>' %}
{% endif %}
{% endif %}
{% if message['role'] == 'user' %}
{% if loop.index0 == 1 and not context.has_head %}
{% set content = '<|startoftext|>' + content %}
{% endif %}
{% if loop.index0 == 1 and context.has_head %}
{% set content = content + '<|extra_0|>' %}
{% else %}
{% set content = '<|startoftext|>' + content + '<|extra_0|>' %}
{% endif %}
{% elif message['role'] == 'assistant' %}
{% set content = content + '<|eos|>' %}
{% elif message['role'] == 'tool' %}
{% set content = content + '<|extra_0|>' %}
{% endif %}
{{- content -}}
{% endfor %}
{% endif %}
{%- if enable_thinking is defined and enable_thinking is false %}
{{- '<think>\n\n</think>\n' }}
{%- endif %}
tests/entrypoints/openai/tool_parsers/test_hunyuan_a13b_tool_parser.py 0 → 100644
View file @ 5a7fb3ab
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# ruff: noqa: E501
import json
from unittest.mock import MagicMock
import pytest
from tests.entrypoints.openai.tool_parsers.utils import (
run_tool_extraction, run_tool_extraction_streaming)
from vllm.entrypoints.openai.protocol import FunctionCall, ToolCall
from vllm.entrypoints.openai.tool_parsers import ToolParser, ToolParserManager
def make_tool_call(name, arguments):
return ToolCall(type="function",
function=FunctionCall(name=name,
arguments=json.dumps(arguments)))
# TODO: add reason prefix and suffix.
@pytest.mark.parametrize(
"model_output,expected_tool_calls,expected_content",
[
# No tool call
("How can I help you today?", [], "How can I help you today?"),
# Single tool call, no content
(
"<tool_calls>[{\"name\": \"get_weather\", \"arguments\": {\"city\": \"San Francisco\", \"metric\": \"celsius\"}}]</tool_calls>", #noqa: E501
[
make_tool_call("get_weather", {
"city": "San Francisco",
"metric": "celsius"
})
],
None),
# Multiple tool calls
(
"<tool_calls>[{\"name\": \"get_weather\", \"arguments\": {\"city\": \"San Francisco\", \"metric\": \"celsius\"}}, {\"name\": \"register_user\", \"arguments\": {\"name\": \"John Doe\", \"age\": 37, \"address\": {\"city\": \"San Francisco\", \"state\": \"CA\"}, \"role\": null, \"passed_test\": true, \"aliases\": [\"John\", \"Johnny\"]}}]</tool_calls>", #noqa: E501
[
make_tool_call("get_weather", {
"city": "San Francisco",
"metric": "celsius"
}),
make_tool_call(
"register_user", {
"name": "John Doe",
"age": 37,
"address": {
"city": "San Francisco",
"state": "CA"
},
"role": None,
"passed_test": True,
"aliases": ["John", "Johnny"]
})
],
None),
# Content before tool call
(
"I will call the tool now. <tool_calls>[{\"name\": \"get_weather\", \"arguments\": {\"city\": \"Boston\"}}]</tool_calls>", #noqa: E501
[make_tool_call("get_weather", {"city": "Boston"})],
"I will call the tool now. "),
# Content after tool call (should be stripped)
(
"<tool_calls>[{\"name\": \"get_weather\", \"arguments\": {\"city\": \"Seattle\"}}]</tool_calls>\nThank you!", #noqa: E501
[make_tool_call("get_weather", {"city": "Seattle"})],
None),
(
"<tool_calls>[{\"name\": \"complex_tool\", \"arguments\": {\"level1\": {\"level2\": {\"level3\": {\"value\": 123}}}}}]</tool_calls>",
[
make_tool_call(
"complex_tool",
{"level1": {
"level2": {
"level3": {
"value": 123
}
}
}})
],
None,
),
])
def test_hunyuan_a13b_tool_parser_extract(model_output, expected_tool_calls,
expected_content):
mock_tokenizer = MagicMock()
tool_parser: ToolParser = ToolParserManager.get_tool_parser(
"hunyuan_a13b")(mock_tokenizer)
content, tool_calls = run_tool_extraction(tool_parser,
model_output,
streaming=False)
# align the random id.
for idx in range(len(tool_calls)):
tool_calls[idx].id = expected_tool_calls[idx].id
assert tool_calls == expected_tool_calls
assert content == expected_content
# Streaming test: simulate incremental output
@pytest.mark.parametrize("model_deltas,expected_tool_calls", [
([
"<tool_calls>[{\"name\": \"get_weather\", ",
"\"arguments\": {\"city\": \"San Francisco\", ",
"\"metric\": \"celsius\"}}]", "</tool_calls>"
], [
make_tool_call("get_weather", {
"city": "San Francisco",
"metric": "celsius"
})
]),
([
"<tool_calls>[{\"name\":", " \"get_weather\",", " \"arguments\":",
" {\"city\": \"Boston\"}", "}]", "</tool_calls>"
], [make_tool_call("get_weather", {"city": "Boston"})]),
([
"", "<tool_calls>[{\"name\":", " \"get_weather\",", " \"arguments\":",
" {\"city\": \"Boston\"}", "}]", "</tool_calls>", "\n</answer>"
], [make_tool_call("get_weather", {"city": "Boston"})]),
pytest.param([
"<tool_calls>[{\"name\": \"complex_tool\",", " \"arguments\": ",
" {\"level1\": {\"level2\": ", "{\"level3\": {\"value\": 123}}}}}",
"]</tool_calls>"
], [
make_tool_call("complex_tool",
{"level1": {
"level2": {
"level3": {
"value": 123
}
}
}})
],
marks=pytest.mark.xfail(
reason="stream parsing not support nested json yet.")),
])
def test_hunyuan_a13b_tool_parser_streaming(model_deltas, expected_tool_calls):
mock_tokenizer = MagicMock()
tool_parser: ToolParser = ToolParserManager.get_tool_parser(
"hunyuan_a13b")(mock_tokenizer)
reconstructor = run_tool_extraction_streaming(
tool_parser, model_deltas, assert_one_tool_per_delta=False)
# align the random id.
for idx in range(len(reconstructor.tool_calls)):
reconstructor.tool_calls[idx].id = expected_tool_calls[idx].id
assert reconstructor.tool_calls == expected_tool_calls
tests/reasoning/test_hunyuan_reasoning_parser.py
View file @ 5a7fb3ab
...@@ -30,6 +30,12 @@ COMPLETE_REASONING = { ...@@ -30,6 +30,12 @@ COMPLETE_REASONING = {
"reasoning_content": "This is a reasoning section", "reasoning_content": "This is a reasoning section",
"content": None, "content": None,
} }
COMPLETE_REASONING_WITH_SYMBOL = {
"output": f"{START_REASONING}This is a reasoning section!{START_RESPONSE}",
"reasoning_content": "This is a reasoning section!",
"content": None,
}
NO_REASONING = { NO_REASONING = {
"output": "This is content", "output": "This is content",
"reasoning_content": None, "reasoning_content": None,
...@@ -70,6 +76,11 @@ TEST_CASES = [ ...@@ -70,6 +76,11 @@ TEST_CASES = [
COMPLETE_REASONING, COMPLETE_REASONING,
id="complete_reasoning", id="complete_reasoning",
), ),
pytest.param(
False,
COMPLETE_REASONING_WITH_SYMBOL,
id="complete_reasoning_with_symbol",
),
pytest.param( pytest.param(
False, False,
NO_REASONING, NO_REASONING,
......
vllm/entrypoints/openai/serving_chat.py
View file @ 5a7fb3ab
...@@ -613,8 +613,13 @@ class OpenAIServingChat(OpenAIServing): ...@@ -613,8 +613,13 @@ class OpenAIServingChat(OpenAIServing):
previous_text = previous_texts[i] previous_text = previous_texts[i]
previous_token_ids = all_previous_token_ids[i] previous_token_ids = all_previous_token_ids[i]
current_text = previous_text + delta_text current_text = previous_text + delta_text
# avoid the None + list error.
if previous_token_ids:
current_token_ids = previous_token_ids + list( current_token_ids = previous_token_ids + list(
output.token_ids) output.token_ids)
else:
current_token_ids = list(output.token_ids)
# handle streaming deltas for tools with named tool_choice # handle streaming deltas for tools with named tool_choice
if tool_choice_function_name: if tool_choice_function_name:
...@@ -1077,9 +1082,17 @@ class OpenAIServingChat(OpenAIServing): ...@@ -1077,9 +1082,17 @@ class OpenAIServingChat(OpenAIServing):
else: else:
# FOR NOW make it a chat message; we will have to detect # FOR NOW make it a chat message; we will have to detect
# the type to make it later. # the type to make it later.
ret_content = content
# try to use content return from tool parser first,
# tool parser may do some modify for the content.
if (tool_call_info.content
and len(tool_call_info.content) > 0):
ret_content = tool_call_info.content
message = ChatMessage(role=role, message = ChatMessage(role=role,
reasoning_content=reasoning_content, reasoning_content=reasoning_content,
content=content) content=ret_content)
# undetermined case that is still important to handle # undetermined case that is still important to handle
else: else:
......
vllm/entrypoints/openai/tool_parsers/__init__.py
View file @ 5a7fb3ab
...@@ -6,6 +6,7 @@ from .deepseekv3_tool_parser import DeepSeekV3ToolParser ...@@ -6,6 +6,7 @@ from .deepseekv3_tool_parser import DeepSeekV3ToolParser
from .granite_20b_fc_tool_parser import Granite20bFCToolParser from .granite_20b_fc_tool_parser import Granite20bFCToolParser
from .granite_tool_parser import GraniteToolParser from .granite_tool_parser import GraniteToolParser
from .hermes_tool_parser import Hermes2ProToolParser from .hermes_tool_parser import Hermes2ProToolParser
from .hunyuan_a13b_tool_parser import HunyuanA13BToolParser
from .internlm2_tool_parser import Internlm2ToolParser from .internlm2_tool_parser import Internlm2ToolParser
from .jamba_tool_parser import JambaToolParser from .jamba_tool_parser import JambaToolParser
from .kimi_k2_tool_parser import KimiK2ToolParser from .kimi_k2_tool_parser import KimiK2ToolParser
...@@ -23,5 +24,5 @@ __all__ = [ ...@@ -23,5 +24,5 @@ __all__ = [
"Internlm2ToolParser", "Llama3JsonToolParser", "JambaToolParser", "Internlm2ToolParser", "Llama3JsonToolParser", "JambaToolParser",
"Llama4PythonicToolParser", "PythonicToolParser", "Phi4MiniJsonToolParser", "Llama4PythonicToolParser", "PythonicToolParser", "Phi4MiniJsonToolParser",
"DeepSeekV3ToolParser", "xLAMToolParser", "MinimaxToolParser", "DeepSeekV3ToolParser", "xLAMToolParser", "MinimaxToolParser",
"KimiK2ToolParser" "KimiK2ToolParser", "HunyuanA13BToolParser"
] ]
vllm/entrypoints/openai/tool_parsers/hunyuan_a13b_tool_parser.py 0 → 100644
View file @ 5a7fb3ab
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# ruff: noqa: E501, SIM102
import json
from collections.abc import Sequence
from typing import Any, Optional, Union
import regex as re
from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
DeltaFunctionCall, DeltaMessage,
DeltaToolCall,
ExtractedToolCallInformation,
FunctionCall, ToolCall)
from vllm.entrypoints.openai.tool_parsers.abstract_tool_parser import (
ToolParser, ToolParserManager)
from vllm.entrypoints.openai.tool_parsers.utils import consume_space
from vllm.logger import init_logger
from vllm.transformers_utils.tokenizer import AnyTokenizer
from vllm.utils import random_uuid
logger = init_logger(__name__)
@ToolParserManager.register_module("hunyuan_a13b")
class HunyuanA13BToolParser(ToolParser):
def __init__(self, tokenizer: AnyTokenizer):
super().__init__(tokenizer)
# Initialize state for streaming mode
self.prev_tool_calls: list[dict] = []
self.current_tool_id = -1
self.current_tool_name_sent = False
self.streamed_args: list[str] = [
] # Track arguments sent for each tool
# For backward compatibility with tests
self.current_tools_sent: list[bool] = []
# For backward compatibility with serving code
self.prev_tool_call_arr = []
# Regex patterns for preprocessing
self.answer_tool_calls_pattern = re.compile(
r"<tool_calls>([\s\S]*?)</tool_calls>", re.DOTALL)
self.tool_name_reg = re.compile(r'"name"\s*:\s*"([^"]+)"')
self.tool_empty_arg_reg = re.compile(
r'"name"\s*:\s*"[^"]+"\s*,\s*"arguments"\s*:\s*\{\s*\}')
# TODO: not support nested json object in fc arguments.
self.tool_non_empty_arg_reg = re.compile(
r'"name"\s*:\s*"[^"]+"\s*,\s*"arguments"\s*:\s*(\{(?:[^{}]|(?:\{[^{}]*\}))*\})'
)
self.bot_string = "<tool_calls>"
# Define streaming state type to be initialized later
self.streaming_state: dict[str, Any] = {
"current_tool_index": -1,
"tool_ids": [],
"sent_tools": [],
}
def preprocess_model_output(
self, model_output: str) -> tuple[Optional[str], Optional[str]]:
# find the location tool call
for match in self.answer_tool_calls_pattern.finditer(model_output):
start, end = match.span()
# check tool_calls whether in side of <think>
think_regions = [(m.start(), m.end()) for m in re.finditer(
r"<think>(.*?)</think>", model_output, flags=re.DOTALL)]
in_think = any(start > t_start and end < t_end
for t_start, t_end in think_regions)
if not in_think:
content = model_output[:start]
tool_calls_content = match.group(1).strip()
try:
json.loads(tool_calls_content)
return content, tool_calls_content
except Exception:
continue
return model_output, None
def extract_tool_calls(
self, model_output: str,
request: ChatCompletionRequest) -> ExtractedToolCallInformation:
"""
Extract tool calls from a complete model output.
"""
try:
# Preprocess the model output
content, potential_tool_calls = self.preprocess_model_output(
model_output)
if not potential_tool_calls:
# some text should be filtered out for no function call
# this text is in a13b's chat template.
if content:
content = content.replace("助手：", "", 1)
return ExtractedToolCallInformation(tools_called=False,
tool_calls=[],
content=content)
# Parse the potential tool calls as JSON
tool_calls_data = json.loads(potential_tool_calls)
# Ensure it's an array
if not isinstance(tool_calls_data, list):
logger.debug("Tool calls data is not an array")
return ExtractedToolCallInformation(
tools_called=False,
tool_calls=[],
content=content or model_output,
)
tool_calls: list[ToolCall] = []
for idx, call in enumerate(tool_calls_data):
if (not isinstance(call, dict) or "name" not in call
or "arguments" not in call):
continue
tool_call = ToolCall(
id=f"call_{random_uuid()}",
type="function",
function=FunctionCall(
name=call["name"],
arguments=(json.dumps(call["arguments"]) if isinstance(
call["arguments"], dict) else call["arguments"]),
),
)
tool_calls.append(tool_call)
if not content or len(content.strip()) == 0:
# clear the whitespace content.
content = None
return ExtractedToolCallInformation(
tools_called=len(tool_calls) > 0,
tool_calls=tool_calls,
content=content,
)
except Exception:
return ExtractedToolCallInformation(tools_called=False,
tool_calls=[],
content=model_output)
def extract_tool_calls_streaming(
self,
previous_text: str,
current_text: str,
delta_text: str,
previous_token_ids: Sequence[int],
current_token_ids: Sequence[int],
delta_token_ids: Sequence[int],
request: ChatCompletionRequest,
) -> Union[DeltaMessage, None]:
"""
Extract tool calls for streaming mode.
"""
start_idx = consume_space(0, current_text)
if current_text[start_idx:].startswith(self.bot_string):
start_idx = consume_space(start_idx + len(self.bot_string),
current_text)
if not current_text or start_idx >= len(
current_text) or current_text[start_idx] != '[':
return DeltaMessage(content=delta_text)
self._try_parse_json_tools(current_text[start_idx:])
test_delta = self._handle_test_compatibility(current_text)
if test_delta:
return test_delta
name_matches = list(self.tool_name_reg.finditer(current_text))
tool_count = len(name_matches)
if tool_count == 0:
return None
self._ensure_state_arrays(tool_count)
current_idx = self.streaming_state["current_tool_index"]
name_delta = self._handle_tool_name_streaming(current_idx, tool_count,
name_matches)
if name_delta:
return name_delta
args_delta = self._handle_tool_args_streaming(current_text,
current_idx, tool_count)
if args_delta:
return args_delta
return None
def _try_parse_json_tools(self, current_text: str):
try:
parsed_tools = json.loads(current_text)
if isinstance(parsed_tools, list):
self.prev_tool_call_arr = parsed_tools
except json.JSONDecodeError:
pass
def _handle_test_compatibility(self, current_text: str):
if len(self.current_tools_sent) > 0:
if (len(self.current_tools_sent) == 1
and self.current_tools_sent[0] is False):
name_match = self.tool_name_reg.search(current_text)
if name_match:
function_name = name_match.group(1)
tool_id = f"chatcmpl-tool-{random_uuid()}"
delta = DeltaMessage(tool_calls=[
DeltaToolCall(
index=0,
type="function",
id=tool_id,
function=DeltaFunctionCall(
name=function_name).model_dump(
exclude_none=True),
)
])
self.current_tools_sent = [True]
self.current_tool_id = 0
self.streaming_state["current_tool_index"] = 0
if len(self.streaming_state["sent_tools"]) == 0:
self.streaming_state["sent_tools"].append({
"sent_name":
True,
"sent_arguments_prefix":
False,
"sent_arguments":
"",
})
else:
self.streaming_state["sent_tools"][0][
"sent_name"] = True
self.current_tool_name_sent = True
return delta
return None
def _ensure_state_arrays(self, tool_count: int):
while len(self.streaming_state["sent_tools"]) < tool_count:
self.streaming_state["sent_tools"].append({
"sent_name": False,
"sent_arguments_prefix": False,
"sent_arguments": "",
})
while len(self.streaming_state["tool_ids"]) < tool_count:
self.streaming_state["tool_ids"].append(None)
def _handle_tool_name_streaming(self, current_idx: int, tool_count: int,
name_matches):
if current_idx == -1 or current_idx < tool_count - 1:
next_idx = current_idx + 1
if (next_idx < tool_count
and not self.streaming_state["sent_tools"][next_idx]
["sent_name"]):
self.streaming_state["current_tool_index"] = next_idx
self.current_tool_id = next_idx
current_idx = next_idx
tool_name = name_matches[current_idx].group(1)
tool_id = f"call_{current_idx}_{random_uuid()}"
self.streaming_state["tool_ids"][current_idx] = tool_id
delta = DeltaMessage(tool_calls=[
DeltaToolCall(
index=current_idx,
type="function",
id=tool_id,
function=DeltaFunctionCall(name=tool_name).model_dump(
exclude_none=True),
)
])
self.streaming_state["sent_tools"][current_idx][
"sent_name"] = True
self.current_tool_name_sent = True
while len(self.streamed_args) <= current_idx:
self.streamed_args.append("")
return delta
return None
def _handle_tool_args_streaming(self, current_text: str, current_idx: int,
tool_count: int):
if current_idx >= 0 and current_idx < tool_count:
empty_args_match = self.tool_empty_arg_reg.search(current_text)
if empty_args_match and empty_args_match.start() > 0:
for i in range(tool_count):
if i == current_idx:
if not self.streaming_state["sent_tools"][current_idx][
"sent_arguments_prefix"]:
self.streaming_state["sent_tools"][current_idx][
"sent_arguments_prefix"] = True
self.streaming_state["sent_tools"][current_idx][
"sent_arguments"] = "{}"
while len(self.streamed_args) <= current_idx:
self.streamed_args.append("")
self.streamed_args[current_idx] += "{}"
delta = DeltaMessage(tool_calls=[
DeltaToolCall(
index=current_idx,
function=DeltaFunctionCall(
arguments="{}").model_dump(
exclude_none=True),
)
])
if current_idx < tool_count - 1:
self.streaming_state["current_tool_index"] += 1
self.current_tool_id = self.streaming_state[
"current_tool_index"]
return delta
args_matches = list(
self.tool_non_empty_arg_reg.finditer(current_text))
if current_idx < len(args_matches):
args_text = args_matches[current_idx].group(1)
is_last_tool = current_idx == tool_count - 1
if not is_last_tool:
next_tool_pos = current_text.find(
"},{", args_matches[current_idx].start())
if next_tool_pos != -1:
args_end_pos = (next_tool_pos + 1)
args_text = (
current_text[args_matches[current_idx].start(
):args_end_pos].split('"arguments":')[1].strip())
sent_args = self.streaming_state["sent_tools"][current_idx][
"sent_arguments"]
if not self.streaming_state["sent_tools"][current_idx][
"sent_arguments_prefix"] and args_text.startswith("{"):
self.streaming_state["sent_tools"][current_idx][
"sent_arguments_prefix"] = True
self.streaming_state["sent_tools"][current_idx][
"sent_arguments"] = "{"
while len(self.streamed_args) <= current_idx:
self.streamed_args.append("")
self.streamed_args[current_idx] += "{"
delta = DeltaMessage(tool_calls=[
DeltaToolCall(
index=current_idx,
function=DeltaFunctionCall(
arguments="{").model_dump(exclude_none=True),
)
])
return delta
if args_text.startswith(sent_args):
args_diff = args_text[len(sent_args):]
if args_diff:
self.streaming_state["sent_tools"][current_idx][
"sent_arguments"] = args_text
while len(self.streamed_args) <= current_idx:
self.streamed_args.append("")
self.streamed_args[current_idx] += args_diff
delta = DeltaMessage(tool_calls=[
DeltaToolCall(
index=current_idx,
function=DeltaFunctionCall(
arguments=args_diff).model_dump(
exclude_none=True),
)
])
return delta
if args_text.endswith("}") and args_text == sent_args:
if current_idx < tool_count - 1:
self.streaming_state["current_tool_index"] += 1
self.current_tool_id = self.streaming_state[
"current_tool_index"]
return None
vllm/model_executor/layers/fused_moe/configs/E=64,N=1536,device_name=NVIDIA_H20,dtype=fp8_w8a8.json 0 → 100644
View file @ 5a7fb3ab
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 5
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 5
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
}
}
vllm/model_executor/layers/fused_moe/configs/E=64,N=3072,device_name=NVIDIA_H20,dtype=fp8_w8a8.json 0 → 100644
View file @ 5a7fb3ab
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 5
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
}
}
vllm/model_executor/layers/fused_moe/configs/E=64,N=3072,device_name=NVIDIA_H20.json 0 → 100644
View file @ 5a7fb3ab
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 5
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
}
}
vllm/model_executor/layers/fused_moe/configs/E=64,N=384,device_name=NVIDIA_H20,dtype=fp8_w8a8.json 0 → 100644
View file @ 5a7fb3ab
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
}
}
vllm/model_executor/layers/fused_moe/configs/E=64,N=384,device_name=NVIDIA_H20.json 0 → 100644
View file @ 5a7fb3ab
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 3
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
}
}
vllm/model_executor/layers/fused_moe/configs/E=64,N=768,device_name=NVIDIA_H20,dtype=fp8_w8a8.json 0 → 100644
View file @ 5a7fb3ab
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
}
}
vllm/model_executor/layers/fused_moe/configs/E=64,N=768,device_name=NVIDIA_H20.json 0 → 100644
View file @ 5a7fb3ab
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
}
}
vllm/reasoning/hunyuan_a13b_reasoning_parser.py
View file @ 5a7fb3ab
...@@ -83,6 +83,13 @@ class HunyuanA13BReasoningParser(ReasoningParser): ...@@ -83,6 +83,13 @@ class HunyuanA13BReasoningParser(ReasoningParser):
def is_reasoning_end(self, input_ids: list[int]) -> bool: def is_reasoning_end(self, input_ids: list[int]) -> bool:
return self.current_state == "response" return self.current_state == "response"
def extract_content_ids(self, input_ids: list[int]) -> list[int]:
# for hunyuan streaming reason parsing, the stream parse
# will call first, and the same token will be called in
# is_reasoning_end and extract_content_ids
# this id is not part of content, so just return [] here.
return []
def extract_reasoning_content( def extract_reasoning_content(
self, model_output: str, request: ChatCompletionRequest self, model_output: str, request: ChatCompletionRequest
) -> tuple[Optional[str], Optional[str]]: ) -> tuple[Optional[str], Optional[str]]:
......
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