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

Add DeepSeek V3.2 support (#3270)

Browse Source 
### What this PR does / why we need it?

This PR added the initial DeepSeek V3.2 support with [vLLM
v0.11.0](https://github.com/vllm-project/vllm/tree/releases/v0.11.0)
(not released yet). We will complete vLLM adaptation as soon as
possible. This feature will be ready in recent 1-2 days.

Related doc: https://github.com/vllm-project/vllm-ascend/pull/3223 .

### Does this PR introduce _any_ user-facing change?
Yes!

### How was this patch tested?
CI passed and Run deepseek doc soon.


- vLLM version: v0.11.0rc3
- vLLM main:
https://github.com/vllm-project/vllm/commit/releases/v0.11.0

---------

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Signed-off-by: zzzzwwjj <1183291235@qq.com>
Signed-off-by: linfeng-yuan <1102311262@qq.com>
Signed-off-by: wxsIcey <1790571317@qq.com>
Signed-off-by: MengqingCao <cmq0113@163.com>
Co-authored-by: zzzzwwjj <1183291235@qq.com>
Co-authored-by: linfeng-yuan <1102311262@qq.com>
Co-authored-by: wxsIcey <1790571317@qq.com>
Co-authored-by: MengqingCao <cmq0113@163.com>
This commit is contained in:
wangxiyuan
2025-09-30 03:25:58 +08:00
committed by GitHub
parent 5503a3142f
commit 81bd6e4c99
27 changed files with 4354 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm_ascend/patch/platform/patch_common/__init__.py
View File

@@ -15,6 +15,10 @@
# limitations under the License.
#
import vllm_ascend.patch.platform.patch_common.patch_config # noqa
import vllm_ascend.patch.platform.patch_common.patch_distributed # noqa
import vllm_ascend.patch.platform.patch_common.patch_mamba_config # noqa
import vllm_ascend.patch.platform.patch_common.patch_multimodal_merge # noqa
import vllm_ascend.patch.platform.patch_common.patch_transformers_utils # noqa
import vllm_ascend.patch.worker.patch_common.patch_attention_selector # noqa
import vllm_ascend.patch.worker.patch_common.patch_attentionspec # noqa

313
vllm_ascend/patch/platform/patch_common/patch_config.py Normal file
View File

@@ -0,0 +1,313 @@
import ast
import vllm.envs as envs
from transformers import PretrainedConfig
from vllm.config import ModelConfig
from vllm.config.speculative import SpeculativeConfig
from vllm.logger import logger
# mypy: ignore-errors
@property
def is_deepseek_mla(self: ModelConfig):
if not hasattr(self.hf_text_config, "model_type"):
return False
elif self.hf_text_config.model_type in \
('deepseek_v2', 'deepseek_v3', 'deepseek_mtp',
'kimi_k2', 'longcat_flash', 'deepseek_v32'):
return self.hf_text_config.kv_lora_rank is not None
elif self.hf_text_config.model_type == 'eagle':
# if the model is an EAGLE module, check for the
# underlying architecture
return self.hf_text_config.model.model_type in \
('deepseek_v2', 'deepseek_v3', 'deepseek_v32') \
and self.hf_text_config.kv_lora_rank is not None
return False
@staticmethod
def hf_config_override(hf_config: PretrainedConfig) -> PretrainedConfig:
if hf_config.model_type in ("deepseek_v3", "deepseek_v32"):
hf_config.model_type = "deepseek_mtp"
if hf_config.model_type == "deepseek_mtp":
n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
hf_config.update({
"n_predict": n_predict,
"architectures": ["DeepSeekMTPModel"]
})
if hf_config.architectures[0] == "MiMoForCausalLM":
hf_config.model_type = "mimo_mtp"
n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
hf_config.update({
"num_hidden_layers": 0,
"n_predict": n_predict,
"architectures": ["MiMoMTPModel"]
})
if hf_config.architectures[0] == "Glm4MoeForCausalLM":
hf_config.model_type = "glm4_moe_mtp"
n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
hf_config.update({
"num_hidden_layers": 0,
"n_predict": n_predict,
"architectures": ["Glm4MoeMTPModel"]
})
if hf_config.model_type == "ernie4_5_moe":
hf_config.model_type = "ernie_mtp"
if hf_config.model_type == "ernie_mtp":
n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
hf_config.update({
"n_predict": n_predict,
"architectures": ["ErnieMTPModel"]
})
if hf_config.model_type == "qwen3_next":
hf_config.model_type = "qwen3_next_mtp"
if hf_config.model_type == "qwen3_next_mtp":
n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
hf_config.update({
"n_predict": n_predict,
"architectures": ["Qwen3NextMTP"]
})
if hf_config.model_type == "longcat_flash":
hf_config.model_type = "longcat_flash_mtp"
n_predict = getattr(hf_config, "num_nextn_predict_layers", 1)
hf_config.update({
"n_predict": n_predict,
"architectures": ["LongCatFlashMTPModel"]
})
return hf_config
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"
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
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"):
if self.enable_chunked_prefill and not envs.VLLM_USE_V1:
raise ValueError(
"Chunked prefill and EAGLE are not compatible "
"when using V0.")
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))
ModelConfig.is_deepseek_mla = is_deepseek_mla
SpeculativeConfig.__post_init__ = __post_init__
SpeculativeConfig.hf_config_override = hf_config_override

5
vllm_ascend/patch/platform/patch_common/patch_mamba_config.py
View File

@@ -6,6 +6,8 @@ from vllm.model_executor.models.config import MambaModelConfig
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, cdiv
from vllm.v1.kv_cache_interface import FullAttentionSpec, MambaSpec
from vllm_ascend.ascend_config import get_ascend_config
@classmethod
def verify_and_update_config(cls, vllm_config) -> None:
@@ -22,6 +24,7 @@ def verify_and_update_config(cls, vllm_config) -> None:
logger = init_logger(__name__)
# Enable FULL_AND_PIECEWISE by default
MambaModelConfig.verify_and_update_config(vllm_config)
ascend_config = get_ascend_config()
cache_config = vllm_config.cache_config
model_config = vllm_config.model_config
@@ -38,7 +41,7 @@ def verify_and_update_config(cls, vllm_config) -> None:
num_kv_heads=model_config.get_num_kv_heads(parallel_config),
head_size=model_config.get_head_size(),
dtype=kv_cache_dtype,
use_mla=model_config.use_mla).page_size_bytes
use_mla=model_config.use_mla or ascend_config.use_sfa).page_size_bytes
model_cls, _ = ModelRegistry.resolve_model_cls(
model_config.architecture,

200
vllm_ascend/patch/platform/patch_common/patch_transformers_utils.py Normal file
View File

@@ -0,0 +1,200 @@
import vllm.transformers_utils.configs
from transformers.configuration_utils import PretrainedConfig
from transformers.utils import logging
from vllm.transformers_utils import config
logger = logging.get_logger(__name__)
class DeepseekV3Config(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`DeepseekV3Model`]. It is used to instantiate an DeepSeek
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
defaults will yield a similar configuration to that of the DeepSeek-V3.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 129280):
Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the
`inputs_ids` passed when calling [`DeepseekV3Model`]
hidden_size (`int`, *optional*, defaults to 4096):
Dimension of the hidden representations.
intermediate_size (`int`, *optional*, defaults to 11008):
Dimension of the MLP representations.
moe_intermediate_size (`int`, *optional*, defaults to 1407):
Dimension of the MoE representations.
num_hidden_layers (`int`, *optional*, defaults to 32):
Number of hidden layers in the Transformer decoder.
num_nextn_predict_layers (`int`, *optional*, defaults to 1):
Number of nextn predict layers in the DeepSeekV3 Model.
num_attention_heads (`int`, *optional*, defaults to 32):
Number of attention heads for each attention layer in the Transformer decoder.
n_shared_experts (`int`, *optional*, defaults to None):
Number of shared experts, None means dense model.
n_routed_experts (`int`, *optional*, defaults to None):
Number of routed experts, None means dense model.
routed_scaling_factor (`float`, *optional*, defaults to 1.0):
Scaling factor or routed experts.
topk_method (`str`, *optional*, defaults to `gready`):
Topk method used in routed gate.
n_group (`int`, *optional*, defaults to None):
Number of groups for routed experts.
topk_group (`int`, *optional*, defaults to None):
Number of selected groups for each token(for each token, ensuring the selected experts is only within `topk_group` groups).
num_experts_per_tok (`int`, *optional*, defaults to None):
Number of selected experts, None means dense model.
moe_layer_freq (`int`, *optional*, defaults to 1):
The frequency of the MoE layer: one expert layer for every `moe_layer_freq - 1` dense layers.
first_k_dense_replace (`int`, *optional*, defaults to 0):
Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head).
\--k dense layers--/
norm_topk_prob (`bool`, *optional*, defaults to False):
Whether to normalize the weights of the routed experts.
scoring_func (`str`, *optional*, defaults to 'softmax'):
Method of computing expert weights.
aux_loss_alpha (`float`, *optional*, defaults to 0.001):
Auxiliary loss weight coefficient.
seq_aux = (`bool`, *optional*, defaults to True):
Whether to compute the auxiliary loss for each individual sample.
num_key_value_heads (`int`, *optional*):
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
`num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
by meanpooling all the original heads within that group. For more details checkout [this
paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
`num_attention_heads`.
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
The non-linear activation function (function or string) in the decoder.
max_position_embeddings (`int`, *optional*, defaults to 2048):
The maximum sequence length that this model might ever be used with.
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
rms_norm_eps (`float`, *optional*, defaults to 1e-06):
The epsilon used by the rms normalization layers.
use_cache (`bool`, *optional*, defaults to `True`):
Whether or not the model should return the last key/values attentions (not used by all models). Only
relevant if `config.is_decoder=True`.
pad_token_id (`int`, *optional*):
Padding token id.
bos_token_id (`int`, *optional*, defaults to 1):
Beginning of stream token id.
eos_token_id (`int`, *optional*, defaults to 2):
End of stream token id.
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
Whether to tie weight embeddings
rope_theta (`float`, *optional*, defaults to 10000.0):
The base period of the RoPE embeddings.
rope_scaling (`Dict`, *optional*):
Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling
strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is
`{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
`max_position_embeddings` to the expected new maximum.
attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
Whether to use a bias in the query, key, value and output projection layers during self-attention.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
```python
>>> from transformers import DeepseekV3Model, DeepseekV3Config
>>> # Initializing a Deepseek-V3 style configuration
>>> configuration = DeepseekV3Config()
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "deepseek_v3"
keys_to_ignore_at_inference = ["past_key_values"]
def __init__(
self,
vocab_size=129280,
hidden_size=7168,
intermediate_size=18432,
moe_intermediate_size=2048,
num_hidden_layers=61,
num_nextn_predict_layers=1,
num_attention_heads=128,
num_key_value_heads=128,
n_shared_experts=1,
n_routed_experts=256,
ep_size=1,
routed_scaling_factor=2.5,
kv_lora_rank=512,
q_lora_rank=1536,
qk_rope_head_dim=64,
v_head_dim=128,
qk_nope_head_dim=128,
topk_method='noaux_tc',
n_group=8,
topk_group=4,
num_experts_per_tok=8,
moe_layer_freq=1,
first_k_dense_replace=3,
norm_topk_prob=True,
scoring_func='sigmoid',
hidden_act="silu",
max_position_embeddings=4096,
initializer_range=0.02,
rms_norm_eps=1e-6,
use_cache=True,
pad_token_id=None,
bos_token_id=0,
eos_token_id=1,
tie_word_embeddings=False,
rope_theta=10000.0,
rope_scaling=None,
attention_bias=False,
attention_dropout=0.0,
**kwargs,
):
self.vocab_size = vocab_size
self.max_position_embeddings = max_position_embeddings
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.moe_intermediate_size = moe_intermediate_size
self.num_hidden_layers = num_hidden_layers
self.num_nextn_predict_layers = num_nextn_predict_layers
self.num_attention_heads = num_attention_heads
self.n_shared_experts = n_shared_experts
self.n_routed_experts = n_routed_experts
self.ep_size = ep_size
self.routed_scaling_factor = routed_scaling_factor
self.kv_lora_rank = kv_lora_rank
self.q_lora_rank = q_lora_rank
self.qk_rope_head_dim = qk_rope_head_dim
self.v_head_dim = v_head_dim
self.qk_nope_head_dim = qk_nope_head_dim
self.topk_method = topk_method
self.n_group = n_group
self.topk_group = topk_group
self.num_experts_per_tok = num_experts_per_tok
self.moe_layer_freq = moe_layer_freq
self.first_k_dense_replace = first_k_dense_replace
self.norm_topk_prob = norm_topk_prob
self.scoring_func = scoring_func
# for backward compatibility
if num_key_value_heads is None:
num_key_value_heads = num_attention_heads
self.num_key_value_heads = num_key_value_heads
self.hidden_act = hidden_act
self.initializer_range = initializer_range
self.rms_norm_eps = rms_norm_eps
self.use_cache = use_cache
self.rope_theta = rope_theta
self.rope_scaling = rope_scaling
self.attention_bias = attention_bias
self.attention_dropout = attention_dropout
super().__init__(
pad_token_id=pad_token_id,
bos_token_id=bos_token_id,
eos_token_id=eos_token_id,
tie_word_embeddings=tie_word_embeddings,
**kwargs,
)
vllm.transformers_utils.configs.__all__.append("DeepseekV3Config")
vllm.transformers_utils.configs.DeepseekV3Config = DeepseekV3Config
config._CONFIG_REGISTRY["deepseek_v32"] = "DeepseekV3Config"