EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

remove useless patch (#4699)

Browse Source 
patach_config is useless now. Let's remove it


- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
This commit is contained in:
wangxiyuan
2025-12-08 11:02:42 +08:00
committed by GitHub
parent 866347a621
commit 0b65ac6c4b
21 changed files with 30 additions and 277 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
examples/external_online_dp/run_dp_template.sh
View File

@@ -29,4 +29,4 @@ vllm serve model_path \
--trust-remote-code \
--gpu-memory-utilization 0.9 \
--quantization ascend \
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}' \
--speculative-config '{"num_speculative_tokens": 1, "method":"mtp"}'

5
tests/e2e/nightly/features/test_mtpx_deepseek_r1_0528_w8a8.py
View File

@@ -74,10 +74,7 @@ async def test_models(model: str, mode: str) -> None:
"VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS": "3600000"
}
additional_config: dict[str, Any] = {}
speculative_config = {
"num_speculative_tokens": 2,
"method": "deepseek_mtp"
}
speculative_config = {"num_speculative_tokens": 2, "method": "mtp"}
compilation_config = {
"cudagraph_capture_sizes": [56],
"cudagraph_mode": "FULL_DECODE_ONLY"

5
tests/e2e/nightly/features/test_prefix_cache_deepseek_r1_0528_w8a8.py
View File

@@ -84,10 +84,7 @@ async def test_models(model: str) -> None:
"chunked_prefill_for_mla": True,
"enable_weight_nz_layout": True
}
speculative_config = {
"num_speculative_tokens": 1,
"method": "deepseek_mtp"
}
speculative_config = {"num_speculative_tokens": 1, "method": "mtp"}
server_args = [
"--quantization", "ascend", "--data-parallel-size", "2",
"--tensor-parallel-size", "8", "--enable-expert-parallel", "--port",

5
tests/e2e/nightly/models/test_deepseek_r1_0528_w8a8.py
View File

@@ -76,10 +76,7 @@ async def test_models(model: str, mode: str) -> None:
"HCCL_BUFFSIZE": "1024",
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True"
}
speculative_config = {
"num_speculative_tokens": 1,
"method": "deepseek_mtp"
}
speculative_config = {"num_speculative_tokens": 1, "method": "mtp"}
additional_config = {
"torchair_graph_config": {
"enabled": True,

5
tests/e2e/nightly/models/test_deepseek_r1_w8a8_eplb.py
View File

@@ -62,10 +62,7 @@ async def test_models(model: str) -> None:
"DISABLE_L2_CACHE": "1",
"DYNAMIC_EPLB": "true",
}
speculative_config = {
"num_speculative_tokens": 1,
"method": "deepseek_mtp"
}
speculative_config = {"num_speculative_tokens": 1, "method": "mtp"}
compilation_config = {
"cudagraph_capture_sizes": [24],
"cudagraph_mode": "FULL_DECODE_ONLY"

4
tests/e2e/nightly/multi_node/config/models/DeepSeek-R1-W8A8-A2-torchair.yaml
View File

@@ -29,7 +29,7 @@ deployment:
--trust-remote-code
--quantization ascend
--gpu-memory-utilization 0.9
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}'
--speculative-config '{"num_speculative_tokens": 1, "method":"mtp"}'
--additional-config '{"torchair_graph_config":{"enabled":true,"enable_multistream_moe":true},"chunked_prefill_for_mla":true,"enable_weight_nz_layout":true}'
-
@@ -50,7 +50,7 @@ deployment:
--trust-remote-code
--quantization ascend
--gpu-memory-utilization 0.9
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}'
--speculative-config '{"num_speculative_tokens": 1, "method":"mtp"}'
--additional-config '{"torchair_graph_config":{"enabled":true,"enable_multistream_moe":true},"chunked_prefill_for_mla":true,"enable_weight_nz_layout":true}'
benchmarks:
acc:

4
tests/e2e/nightly/multi_node/config/models/DeepSeek-R1-W8A8-A2.yaml
View File

@@ -30,7 +30,7 @@ deployment:
--quantization ascend
--gpu-memory-utilization 0.9
--enforce-eager
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}'
--speculative-config '{"num_speculative_tokens": 1, "method":"mtp"}'
--additional-config '{"chunked_prefill_for_mla":true,"enable_weight_nz_layout":true}'
-
@@ -52,6 +52,6 @@ deployment:
--quantization ascend
--gpu-memory-utilization 0.9
--enforce-eager
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}'
--speculative-config '{"num_speculative_tokens": 1, "method":"mtp"}'
--additional-config '{"chunked_prefill_for_mla":true,"enable_weight_nz_layout":true}'
benchmarks:

8
tests/e2e/nightly/multi_node/config/models/DeepSeek-R1-W8A8-EPLB.yaml
View File

@@ -39,7 +39,7 @@ deployment:
--max-num-batched-tokens 16384
--trust-remote-code
--gpu-memory-utilization 0.9
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}'
--speculative-config '{"num_speculative_tokens": 1, "method":"mtp"}'
--kv-transfer-config
'{"kv_connector": "LLMDataDistCMgrConnector",
"kv_buffer_device": "npu",
@@ -69,7 +69,7 @@ deployment:
--max-num-batched-tokens 16384
--trust-remote-code
--gpu-memory-utilization 0.9
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}'
--speculative-config '{"num_speculative_tokens": 1, "method":"mtp"}'
--kv-transfer-config
'{"kv_connector": "LLMDataDistCMgrConnector",
"kv_buffer_device": "npu",
@@ -100,7 +100,7 @@ deployment:
--max-num-batched-tokens 256
--trust-remote-code
--gpu-memory-utilization 0.9
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}'
--speculative-config '{"num_speculative_tokens": 1, "method":"mtp"}'
--kv-transfer-config
'{"kv_connector": "LLMDataDistCMgrConnector",
"kv_buffer_device": "npu",
@@ -130,7 +130,7 @@ deployment:
--max-num-batched-tokens 256
--trust-remote-code
--gpu-memory-utilization 0.9
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}'
--speculative-config '{"num_speculative_tokens": 1, "method":"mtp"}'
--kv-transfer-config
'{"kv_connector": "LLMDataDistCMgrConnector",
"kv_buffer_device": "npu",

8
tests/e2e/nightly/multi_node/config/models/DeepSeek-R1-W8A8.yaml
View File

@@ -38,7 +38,7 @@ deployment:
--max-num-batched-tokens 16384
--trust-remote-code
--gpu-memory-utilization 0.9
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}'
--speculative-config '{"num_speculative_tokens": 1, "method":"mtp"}'
--kv-transfer-config
'{"kv_connector": "LLMDataDistCMgrConnector",
"kv_buffer_device": "npu",
@@ -68,7 +68,7 @@ deployment:
--max-num-batched-tokens 16384
--trust-remote-code
--gpu-memory-utilization 0.9
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}'
--speculative-config '{"num_speculative_tokens": 1, "method":"mtp"}'
--kv-transfer-config
'{"kv_connector": "LLMDataDistCMgrConnector",
"kv_buffer_device": "npu",
@@ -99,7 +99,7 @@ deployment:
--max-num-batched-tokens 256
--trust-remote-code
--gpu-memory-utilization 0.9
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}'
--speculative-config '{"num_speculative_tokens": 1, "method":"mtp"}'
--kv-transfer-config
'{"kv_connector": "LLMDataDistCMgrConnector",
"kv_buffer_device": "npu",
@@ -129,7 +129,7 @@ deployment:
--max-num-batched-tokens 256
--trust-remote-code
--gpu-memory-utilization 0.9
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}'
--speculative-config '{"num_speculative_tokens": 1, "method":"mtp"}'
--kv-transfer-config
'{"kv_connector": "LLMDataDistCMgrConnector",
"kv_buffer_device": "npu",

2
tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_correctness.py
View File

@@ -56,7 +56,7 @@ def mtp_correctness(sampling_config: SamplingParams,
enable_expert_parallel=True,
speculative_config={
"method":
"deepseek_mtp",
"mtp",
"num_speculative_tokens":
num_speculative_tokens,
"disable_padded_drafter_batch":

2
tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_torchair_correctness.py
View File

@@ -58,7 +58,7 @@ def mtp_torchair_correctness(
distributed_executor_backend="mp",
enable_expert_parallel=True,
speculative_config={
"method": "deepseek_mtp",
"method": "mtp",
"num_speculative_tokens": 1,
},
enforce_eager=False,

2
tests/ut/torchair/test_torchair_model_runner.py
View File

@@ -21,7 +21,7 @@ class TestNPUTorchairModelRunner(PytestBase):
runner.vllm_config = MagicMock(spec=VllmConfig)
runner.speculative_config = MagicMock(
method="deepseek_mtp",
method="mtp",
num_speculative_tokens=4,
disable_padded_drafter_batch=False)

2
tests/ut/torchair/test_torchair_mtp_proposer.py
View File

@@ -19,7 +19,7 @@ class TestTorchairMtpProposer(PytestBase):
vllm_config.speculative_config = MagicMock()
vllm_config.speculative_config.draft_model_config = MagicMock()
vllm_config.speculative_config.draft_model_config.dtype = torch.float16
vllm_config.speculative_config.method = "deepseek_mtp"
vllm_config.speculative_config.method = "mtp"
vllm_config.speculative_config.num_speculative_tokens = 5
vllm_config.load_config = MagicMock()
cache_config = CacheConfig(block_size=16)

2
vllm_ascend/compilation/acl_graph.py
View File

@@ -257,7 +257,7 @@ def update_mla_attn_params(update_stream, forward_context, runtime_shape,
softmax_lse) = param
seq_lens_list = forward_context.attn_metadata[
key].decode.seq_lens_list
if speculative_config and speculative_config.method == "deepseek_mtp" \
if speculative_config and speculative_config.method == "mtp" \
and not forward_context.is_mtp_model:
actual_seq_lengths = forward_context.attn_metadata[
key].decode.actual_seq_lengths_q

1
vllm_ascend/patch/platform/__init__.py
View File

@@ -16,7 +16,6 @@
import os
import vllm_ascend.patch.platform.patch_config # noqa
import vllm_ascend.patch.platform.patch_distributed # noqa
import vllm_ascend.patch.platform.patch_ec_connector # noqa
import vllm_ascend.patch.platform.patch_mamba_config # noqa

234
vllm_ascend/patch/platform/patch_config.py
View File

@@ -1,234 +0,0 @@
import ast
from vllm.config.speculative import SpeculativeConfig
from vllm.logger import logger
def __post_init__(self):
# Note: "method" is a new parameter that helps to extend the
# configuration of non-model-based proposers, and the "model" parameter
# will be used to set the draft model, eagle head, or additional weight
# when needed. If users do not specify "method", the speculative method
# will be detected automatically if possible. If the speculative method
# can not be detected, it will be considered as the "draft_model" by
# default.
if self.model is None and self.num_speculative_tokens is not None:
# TODO(Shangming): Refactor mtp configuration logic when supporting
if (self.target_model_config
and self.target_model_config.hf_text_config.model_type
in ("deepseek_v3", "deepseek_v32", "mimo", "ernie4_5_moe",
"qwen3_next")):
# use the draft model from the same model:
self.model = self.target_model_config.model
# Align the quantization of draft model for cases such as
# --quantization fp8 with a bf16 checkpoint.
if not self.quantization:
self.quantization = self.target_model_config.quantization
elif self.method in ("ngram", "[ngram]"):
self.model = "ngram"
elif self.method == "suffix":
self.model = "suffix"
else:
raise ValueError("num_speculative_tokens was provided but without "
"speculative model.")
# Automatically configure the method for ngram when "model" is used
# instead of "method"
if self.method is None and (self.model is not None
and self.model in ("ngram", "[ngram]")):
self.method = "ngram"
if self.method in ("ngram", "[ngram]"):
# Unified to "ngram" internally
self.method = "ngram"
# Set default values if not provided
if (self.prompt_lookup_min is None and self.prompt_lookup_max is None):
# TODO(woosuk): Tune these values. They are arbitrarily chosen.
self.prompt_lookup_min = 5
self.prompt_lookup_max = 5
elif self.prompt_lookup_min is None:
assert self.prompt_lookup_max is not None
self.prompt_lookup_min = self.prompt_lookup_max
elif self.prompt_lookup_max is None:
assert self.prompt_lookup_min is not None
self.prompt_lookup_max = self.prompt_lookup_min
# Validate values
if self.prompt_lookup_min < 1:
raise ValueError(
f"prompt_lookup_min={self.prompt_lookup_min} must be > 0")
if self.prompt_lookup_max < 1:
raise ValueError(
f"prompt_lookup_max={self.prompt_lookup_max} must be > 0")
if self.prompt_lookup_min > self.prompt_lookup_max:
raise ValueError(
f"prompt_lookup_min={self.prompt_lookup_min} must "
f"be <= prompt_lookup_max={self.prompt_lookup_max}")
# TODO: current we still need extract vocab_size from target model
# config, in future, we may try refactor it out, and set
# draft related config as None here.
self.draft_model_config = self.target_model_config
self.draft_parallel_config = self.target_parallel_config
elif self.method == "suffix":
self.draft_model_config = self.target_model_config
self.draft_parallel_config = self.target_parallel_config
self._validate_suffix_decoding()
else:
self.prompt_lookup_max = 0
self.prompt_lookup_min = 0
if self.model is not None:
# TODO: Move this import to the top once `ModelConfig`
# lives in `vllm.config.model`.
from vllm.config import ModelConfig
self.draft_model_config = ModelConfig(
model=self.model,
runner="draft",
tokenizer=self.target_model_config.tokenizer,
tokenizer_mode=self.target_model_config.tokenizer_mode,
trust_remote_code=self.target_model_config.trust_remote_code,
allowed_local_media_path=self.target_model_config.
allowed_local_media_path,
allowed_media_domains=self.target_model_config.
allowed_media_domains,
dtype=self.target_model_config.dtype,
seed=self.target_model_config.seed,
revision=self.revision,
code_revision=self.code_revision,
tokenizer_revision=self.target_model_config.tokenizer_revision,
spec_target_max_model_len=self.target_model_config.
max_model_len,
quantization=self.quantization,
enforce_eager=self.target_model_config.enforce_eager,
max_logprobs=self.target_model_config.max_logprobs,
hf_overrides=SpeculativeConfig.hf_config_override,
)
# Automatically detect the method
if self.method in ('eagle', 'eagle3'):
pass
# examples:
# yuhuili/EAGLE-LLaMA3-Instruct-8B
# yuhuili/EAGLE3-LLaMA3.1-Instruct-8B
# AngelSlim/Qwen3-8B_eagle3
elif "eagle-" in self.draft_model_config.model.lower():
self.method = "eagle"
elif "eagle3" in self.draft_model_config.model.lower():
self.method = "eagle3"
elif self.draft_model_config.hf_config.model_type == "medusa":
self.method = "medusa"
elif (self.draft_model_config.hf_config.model_type ==
"mlp_speculator"):
self.method = "mlp_speculator"
elif (self.draft_model_config.hf_config.model_type
in ("deepseek_mtp", "mimo_mtp", "glm4_moe_mtp")):
self.method = "deepseek_mtp"
if self.num_speculative_tokens > 1:
logger.warning(
"All Deepseek MTP models only have " \
"one layer. Might need some code changes " \
"to support multiple layers."
)
elif (self.draft_model_config.hf_config.model_type == "ernie_mtp"):
self.method = "ernie_mtp"
if self.num_speculative_tokens > 1:
logger.warning(
"All Ernie MTP models only have " \
"one layer. Might need some code changes " \
"to support multiple layers."
)
elif (self.draft_model_config.hf_config.model_type ==
"qwen3_next_mtp"):
self.method = "qwen3_next_mtp"
if self.num_speculative_tokens > 1:
logger.warning(
"All Qwen3Next MTP models only have " \
"one layer. Might need some code changes " \
"to support multiple layers."
)
elif (self.draft_model_config.hf_config.model_type
in ("longcat_flash_mtp")):
self.method = "longcat_flash_mtp"
if self.num_speculative_tokens > 1:
logger.warning(
"LongCat MTP models only have " \
"one layer. Might need some code changes " \
"to support multiple layers."
)
else:
self.method = "draft_model"
raise NotImplementedError(
"Speculative decoding with draft model is not "
"supported yet. Please consider using other "
"speculative decoding methods such as ngram, medusa, "
"eagle, or deepseek_mtp.")
# Replace hf_config for EAGLE draft_model
if self.method in ("eagle", "eagle3"):
from vllm.transformers_utils.configs import SpeculatorsConfig
from vllm.transformers_utils.configs.eagle import EAGLEConfig
if isinstance(self.draft_model_config.hf_config,
(EAGLEConfig, SpeculatorsConfig)):
pass
else:
eagle_config = EAGLEConfig(
self.draft_model_config.hf_config,
method=self.method,
model_type="eagle")
self.draft_model_config.hf_config = eagle_config
if (self.num_speculative_tokens is not None
and hasattr(self.draft_model_config.hf_config,
"num_lookahead_tokens")):
self.draft_model_config.hf_config.num_lookahead_tokens = \
self.num_speculative_tokens
n_predict = getattr(self.draft_model_config.hf_config, "n_predict",
None)
if n_predict is not None:
if self.num_speculative_tokens is None:
# Default to max value defined in draft model config.
self.num_speculative_tokens = n_predict
elif self.num_speculative_tokens > n_predict and \
self.num_speculative_tokens % n_predict != 0:
# Ensure divisibility for MTP module reuse.
raise ValueError(
f"num_speculative_tokens:{self.num_speculative_tokens}"
f" must be divisible by {n_predict=}")
if self.speculative_token_tree is None:
# Generate chain of tokens.
self.speculative_token_tree = str([
(i + 1) * (0, ) for i in range(self.num_speculative_tokens)
])
else:
# Sort the token tree breadth-first.
tree_choices = ast.literal_eval(self.speculative_token_tree)
self.speculative_token_tree = str(
sorted(tree_choices, key=lambda t: (len(t), t)))
self.draft_tensor_parallel_size = \
SpeculativeConfig._verify_and_get_draft_tp(
self.target_parallel_config,
self.draft_tensor_parallel_size,
self.draft_model_config.hf_config
)
self.draft_model_config.max_model_len = (
SpeculativeConfig._maybe_override_draft_max_model_len(
self.max_model_len,
self.draft_model_config.max_model_len,
self.target_model_config.max_model_len,
))
self.draft_parallel_config = (
SpeculativeConfig.create_draft_parallel_config(
self.target_parallel_config,
self.draft_tensor_parallel_size))
SpeculativeConfig.__post_init__ = __post_init__

2
vllm_ascend/spec_decode/__init__.py
View File

@@ -32,7 +32,7 @@ def get_spec_decode_method(method,
return NgramProposer(vllm_config, device, runner)
elif method in ("eagle", "eagle3"):
return EagleProposer(vllm_config, device, runner)
elif method in ('deepseek_mtp', 'qwen3_next_mtp'):
elif method == "mtp":
if is_torchair_graph:
return TorchairMtpProposer(vllm_config, device, runner)
return MtpProposer(vllm_config, device, runner)

2
vllm_ascend/torchair/torchair_mla.py
View File

@@ -317,7 +317,7 @@ class AscendMLATorchairMetadataBuilder:
dtype=self.model_config.dtype,
device=device)
if self.vllm_config.speculative_config is not None and\
self.vllm_config.speculative_config.method == 'deepseek_mtp':
self.vllm_config.speculative_config.method == 'mtp':
attn_state = AscendAttentionState.SpecDecoding
num_decode_tokens = 2
else:

2
vllm_ascend/torchair/torchair_model_runner.py
View File

@@ -501,7 +501,7 @@ class NPUTorchairModelRunner(NPUModelRunner):
def update_torchair_graph_batch_sizes(self):
# return graph_batch_sizes according to the max number of tokens
# first pad according to the number of requests
if self.is_kv_consumer and self.speculative_config and self.speculative_config.method == 'deepseek_mtp':
if self.is_kv_consumer and self.speculative_config and self.speculative_config.method == 'mtp':
# pd disaggregation scenario may incorrectly calculate the batch in mtp scenario, so we force set it to max_num_reqs
self.torchair_graph_batch_sizes = [self.max_num_reqs]
logger.warning(

2
vllm_ascend/torchair/torchair_sfa.py
View File

@@ -319,7 +319,7 @@ class AscendSFATorchairMetadataBuilder:
device=device)
if self.vllm_config.speculative_config is not None and\
self.vllm_config.speculative_config.method == 'deepseek_mtp':
self.vllm_config.speculative_config.method == 'mtp':
attn_state = AscendAttentionState.SpecDecoding
num_decode_tokens = 2
else:

8
vllm_ascend/worker/model_runner_v1.py
View File

@@ -2044,13 +2044,13 @@ class NPUModelRunner(LoRAModelRunnerMixin, ECConnectorModelRunnerMixin):
# We assume it is the decode stage, where prefill occurs but only one token is not hit in cache.
elif np.all(num_scheduled_tokens == 1):
attn_state = AscendAttentionState.DecodeOnly
if self.speculative_config and self.speculative_config.method == 'deepseek_mtp':
if self.speculative_config and self.speculative_config.method == 'mtp':
# SpecDecoding now supports seq_len=1 and seq_len=2
# In Prefilling Decoding Disaggregation scenario, SpecDecoding need to supports seq_len=1
attn_state = AscendAttentionState.SpecDecoding
# Speculative decoding.
elif np.all(num_valid_tokens == 1):
if self.speculative_config and self.speculative_config.method == 'deepseek_mtp':
if self.speculative_config and self.speculative_config.method == 'mtp':
attn_state = AscendAttentionState.SpecDecoding
else:
attn_state = AscendAttentionState.ChunkedPrefill
@@ -2701,7 +2701,7 @@ class NPUModelRunner(LoRAModelRunnerMixin, ECConnectorModelRunnerMixin):
with ProfileExecuteDuration().capture_async("Draft"):
if self.speculative_config:
use_padded_batch_for_eagle = self.speculative_config and \
self.speculative_config.method in ("deepseek_mtp", "qwen3_next_mtp") and \
self.speculative_config.method == "mtp" and \
not self.speculative_config.disable_padded_drafter_batch
if use_padded_batch_for_eagle:
# EAGLE speculative decoding can use the GPU sampled tokens
@@ -2900,7 +2900,7 @@ class NPUModelRunner(LoRAModelRunnerMixin, ECConnectorModelRunnerMixin):
block_table_tensor[:num_reqs * self.decode_threshold]
attn_state = AscendAttentionState.DecodeOnly
if self.speculative_config and \
self.speculative_config.method == "deepseek_mtp":
self.speculative_config.method == "mtp":
attn_state = AscendAttentionState.SpecDecoding
common_metadata = CommonAttentionMetadata(