[Model] Support Qwen3.8-Flash-Next (#53896)

Signed-off-by: huanghaoyan.hhy <huanghaoyan.hhy@alibaba-inc.com>
Signed-off-by: andyluo7 <andy.luo@amd.com>
Signed-off-by: lishunyang12 <lishunyang03@gmail.com>
Signed-off-by: Harry Huang <huanghaoyan.hhy@alibaba-inc.com>
Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
Signed-off-by: Isotr0py <Isotr0py@outlook.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Yifan Qiao <yifanqiao@inferact.ai>
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Co-authored-by: zjy0516 <riverclouds.zhu@qq.com>
Co-authored-by: KungYork <firelikest@163.com>
Co-authored-by: Thien Tran <gau.nernst@yahoo.com.sg>
Co-authored-by: Yifan Qiao <yifanqiao@inferact.ai>
Co-authored-by: andyluo7 <43718156+andyluo7@users.noreply.github.com>
Co-authored-by: Isotr0py <mozf@inferact.ai>
Co-authored-by: andyluo7 <andy.luo@amd.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
Co-authored-by: lishunyang12 <lishunyang03@gmail.com>
Co-authored-by: Vadim Gimpelson <156319763+vadiklyutiy@users.noreply.github.com>
Co-authored-by: Isotr0py <Isotr0py@outlook.com>
Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Harry Huang
2026-08-31 05:57:56 +00:00
committed by GitHub
co-authored by zjy0516 KungYork Thien Tran Yifan Qiao andyluo7 Isotr0py andyluo7 OpenAI Codex lishunyang12 Vadim Gimpelson Isotr0py Lucas Wilkinson Claude Opus 5 OpenAI Codex
parent 687db59744
commit e126687a9a
124 changed files with 19424 additions and 450 deletions
+80
View File
@@ -839,6 +839,83 @@ class Qwen3_5ForCausalLMConfig(Qwen3_5ForConditionalGenerationConfig):
rope_parameters.pop("mrope_interleaved", None)
def _strip_qwen4_exp_mrope(model_config: "ModelConfig") -> None:
configs = {
id(config): config
for config in (
getattr(model_config, "hf_config", None),
model_config.hf_text_config,
)
if config is not None
}
for config in configs.values():
rope_parameters = getattr(config, "rope_parameters", None)
if rope_parameters is not None:
rope_parameters.pop("mrope_section", None)
rope_parameters.pop("mrope_interleaved", None)
class Qwen4ExpForConditionalGenerationConfig(Qwen3_5ForConditionalGenerationConfig):
"""Apply the Qwen3.5 hybrid-cache contract to Qwen4Exp."""
@staticmethod
def verify_and_update_config(vllm_config: "VllmConfig") -> None:
Qwen3_5ForConditionalGenerationConfig.verify_and_update_config(vllm_config)
text_config = vllm_config.model_config.hf_text_config
if text_config.hc_count <= 1:
raise ValueError("Qwen4Exp requires hc_count > 1")
parallel_config = vllm_config.parallel_config
uses_ple_or_qsa = bool(text_config.ple_layer_ids) or (
getattr(text_config, "indexer_n_heads", None) is not None
)
if uses_ple_or_qsa and (
parallel_config.enable_dbo or parallel_config.ubatch_size > 1
):
raise NotImplementedError(
"Qwen4Exp PLE/QSA does not support dual-batch overlap or microbatching"
)
# Checked again in Qwen4ExpModelState; rejecting it here keeps the
# engine from loading weights first.
if text_config.ple_layer_ids and parallel_config.pipeline_parallel_size > 1:
raise NotImplementedError(
"Qwen4Exp N-gram PLE embedding requires pipeline_parallel_size=1 "
"because non-first pipeline ranks do not receive the raw input_ids "
"it needs. Please run with PP=1."
)
multimodal_config = vllm_config.model_config.multimodal_config
if multimodal_config is not None and multimodal_config.language_model_only:
_strip_qwen4_exp_mrope(vllm_config.model_config)
spec_config = vllm_config.speculative_config
if spec_config is not None and spec_config.method not in {
"mtp",
"ngram",
"ngram_gpu",
}:
raise NotImplementedError(
"Qwen4Exp speculative decoding supports only its native MTP "
"checkpoint and linear n-gram proposers"
)
class Qwen4ExpForCausalLMConfig(Qwen4ExpForConditionalGenerationConfig):
@staticmethod
def verify_and_update_config(vllm_config: "VllmConfig") -> None:
Qwen4ExpForConditionalGenerationConfig.verify_and_update_config(vllm_config)
_strip_qwen4_exp_mrope(vllm_config.model_config)
class Qwen4ExpMTPConfig(Qwen4ExpForConditionalGenerationConfig):
"""Preserve MRoPE for a VL target and use 1D RoPE for a text target."""
@staticmethod
def verify_and_update_config(vllm_config: "VllmConfig") -> None:
Qwen4ExpForConditionalGenerationConfig.verify_and_update_config(vllm_config)
if hasattr(vllm_config.model_config.hf_config, "vision_config"):
return
_strip_qwen4_exp_mrope(vllm_config.model_config)
class ColQwen3_5Config(Qwen3_5ForConditionalGenerationConfig):
"""Apply the attention contract declared by a ColQwen3.5 checkpoint."""
@@ -963,6 +1040,9 @@ MODELS_CONFIG_MAP: dict[str, type[VerifyAndUpdateConfig]] = {
"Qwen3_5ForConditionalGeneration": Qwen3_5ForConditionalGenerationConfig,
"Qwen3_5MoeForCausalLM": Qwen3_5ForCausalLMConfig,
"Qwen3_5MoeForConditionalGeneration": Qwen3_5ForConditionalGenerationConfig,
"Qwen4ExpForCausalLM": Qwen4ExpForCausalLMConfig,
"Qwen4ExpForConditionalGeneration": (Qwen4ExpForConditionalGenerationConfig),
"Qwen4ExpMTP": Qwen4ExpMTPConfig,
"UnlimitedOCRForCausalLM": UnlimitedOCRForCausalLMConfig,
"VoyageQwen3BidirectionalEmbedModel": VoyageQwen3BidirectionalEmbedModelConfig,
"XLMRobertaModel": JinaRobertaModelConfig,
+27 -1
View File
@@ -49,13 +49,17 @@ if TYPE_CHECKING:
from vllm.lora.model_manager import LoRAModelManager
from vllm.model_executor.layers.fused_moe import MoERunner
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.mamba.mamba_utils import MambaStateCopyFunc
from vllm.model_executor.layers.mamba.mamba_utils import (
MambaStateCopyFunc,
MambaStateCopyFuncsByType,
)
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
from vllm.model_executor.models.utils import WeightsMapper
from vllm.multimodal.inputs import MultiModalFeatureSpec, MultiModalKwargsItem
from vllm.multimodal.registry import _ProcessorFactories
from vllm.sequence import IntermediateTensors
from vllm.tasks import ScoreType
from vllm.v1.attention.backends.registry import MambaAttentionBackendEnum
from vllm.v1.worker.encoder_cudagraph_defs import (
EncoderCudaGraphCaptureInputs,
EncoderCudaGraphConfig,
@@ -970,6 +974,14 @@ class IsHybrid(Protocol):
"""
...
@classmethod
def get_mamba_state_copy_funcs(
cls,
mamba_types: set["MambaAttentionBackendEnum"],
) -> "MambaStateCopyFuncsByType":
copy_funcs = cls.get_mamba_state_copy_func()
return {mamba_type: copy_funcs for mamba_type in mamba_types}
@overload
def is_hybrid(model: type[object]) -> TypeIs[type[IsHybrid]]: ...
@@ -1131,6 +1143,20 @@ class SupportsMambaPrefixCaching(Protocol):
supports_mamba_prefix_caching: ClassVar[Literal[True]] = True
@classmethod
def get_mamba_state_copy_func(cls) -> tuple["MambaStateCopyFunc", ...]:
"""Return copy functions for the model's Mamba states."""
...
@classmethod
def get_mamba_state_copy_funcs(
cls,
mamba_types: set["MambaAttentionBackendEnum"],
) -> "MambaStateCopyFuncsByType":
"""Map legacy copy functions to each requested Mamba backend."""
copy_funcs = cls.get_mamba_state_copy_func()
return {mamba_type: copy_funcs for mamba_type in mamba_types}
@overload
def supports_mamba_prefix_caching(
+4 -2
View File
@@ -80,15 +80,17 @@ class Qwen2MoeMLP(nn.Module):
reduce_results: bool = True,
expert_gate: torch.nn.Linear | None = None,
is_sequence_parallel: bool = False,
disable_tp: bool = False,
prefix: str = "",
) -> None:
super().__init__()
disable_tp = disable_tp or is_sequence_parallel
self.gate_up_proj = MergedColumnParallelLinear(
hidden_size,
[intermediate_size] * 2,
bias=False,
quant_config=quant_config,
disable_tp=is_sequence_parallel,
disable_tp=disable_tp,
prefix=f"{prefix}.gate_up_proj",
)
self.down_proj = RowParallelLinear(
@@ -97,7 +99,7 @@ class Qwen2MoeMLP(nn.Module):
bias=False,
quant_config=quant_config,
reduce_results=reduce_results,
disable_tp=is_sequence_parallel,
disable_tp=disable_tp,
prefix=f"{prefix}.down_proj",
)
if hidden_act != "silu":
+62 -3
View File
@@ -43,6 +43,9 @@ from vllm.model_executor.layers.mamba.mamba_utils import (
MambaStateShapeCalculator,
)
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.quantization.utils.config_utils import (
get_quark_ocp_mx_group_size,
)
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding, get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
ParallelLMHead,
@@ -90,6 +93,38 @@ def _should_use_sequence_parallel(vllm_config: VllmConfig) -> bool:
)
def _should_replicate_misaligned_shared_expert(
intermediate_size: int,
tp_size: int,
group_size: int | None,
enable_expert_parallel: bool,
is_sequence_parallel: bool,
) -> bool:
if intermediate_size <= 0 or group_size is None:
return False
partition_size, remainder = divmod(intermediate_size, tp_size)
if remainder == 0 and partition_size % group_size == 0:
return False
if enable_expert_parallel or is_sequence_parallel:
return True
if remainder != 0:
raise ValueError(
f"Shared-expert intermediate size {intermediate_size} must be "
f"divisible by tensor-parallel size {tp_size}."
)
raise ValueError(
"The Quark OCP MX shared expert cannot be tensor-parallelized: "
f"intermediate size {intermediate_size} with TP size {tp_size} "
f"produces a partition of {partition_size}, which is not divisible by "
f"the OCP MX group size {group_size}. Choose a compatible "
"tensor-parallel size or enable expert parallelism."
)
class Qwen3NextSparseMoeBlock(nn.Module):
def __init__(self, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -106,6 +141,17 @@ class Qwen3NextSparseMoeBlock(nn.Module):
self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe
shared_expert_group_size = get_quark_ocp_mx_group_size(
quant_config,
f"{prefix}.shared_expert.down_proj",
)
self.replicate_shared_expert = _should_replicate_misaligned_shared_expert(
config.shared_expert_intermediate_size,
self.tp_size,
shared_expert_group_size,
parallel_config.enable_expert_parallel,
self.is_sequence_parallel,
)
if self.tp_size > config.num_experts:
raise ValueError(
f"Tensor parallel size {self.tp_size} is greater than "
@@ -138,7 +184,10 @@ class Qwen3NextSparseMoeBlock(nn.Module):
)
self.is_fused_shared_expert_enabled = False
if config.shared_expert_intermediate_size > 0:
if (
config.shared_expert_intermediate_size > 0
and not self.replicate_shared_expert
):
self.is_fused_shared_expert_enabled = resolve_layer_fused_shared_expert(
quant_config,
prefix,
@@ -159,11 +208,14 @@ class Qwen3NextSparseMoeBlock(nn.Module):
reduce_results=False,
expert_gate=self.shared_expert_gate,
is_sequence_parallel=self.is_sequence_parallel,
disable_tp=self.replicate_shared_expert,
prefix=f"{prefix}.shared_expert",
)
self.experts = FusedMoEFactory(
shared_experts=self.shared_expert,
shared_experts=(
None if self.replicate_shared_expert else self.shared_expert
),
gate=self.gate,
num_experts=self.n_routed_experts,
top_k=config.num_experts_per_tok,
@@ -195,9 +247,16 @@ class Qwen3NextSparseMoeBlock(nn.Module):
if self.is_sequence_parallel and not already_sequence_parallel:
hidden_states = sequence_parallel_chunk(hidden_states)
replicated_shared_output = (
self.shared_expert(hidden_states)
if self.replicate_shared_expert and self.shared_expert is not None
else None
)
final_hidden_states = self.experts(
hidden_states=hidden_states, router_logits=hidden_states
)
if replicated_shared_output is not None:
final_hidden_states += replicated_shared_output
if self.is_sequence_parallel and not already_sequence_parallel:
final_hidden_states = tensor_model_parallel_all_gather(
@@ -488,7 +547,7 @@ class Qwen3NextDecoderLayer(nn.Module):
self,
hidden_states: torch.Tensor,
residual: torch.Tensor | None,
positions: torch.Tensor = None,
positions: torch.Tensor,
**kwargs: object,
):
full_num_tokens = positions.shape[-1]
+9
View File
@@ -111,6 +111,10 @@ _TEXT_GENERATION_MODELS = {
"Gemma3nForCausalLM": ("gemma3n", "Gemma3nForCausalLM"),
"Gemma4ForCausalLM": ("gemma4", "Gemma4ForCausalLM"),
"Qwen3NextForCausalLM": ("qwen3_next", "Qwen3NextForCausalLM"),
"Qwen4ExpForCausalLM": (
"vllm.models.qwen4_exp",
"Qwen4ExpForCausalLM",
),
"GlmForCausalLM": ("glm", "GlmForCausalLM"),
"Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
"Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
@@ -573,6 +577,10 @@ _MULTIMODAL_MODELS = {
"qwen3_5",
"Qwen3_5MoeForConditionalGeneration",
),
"Qwen4ExpForConditionalGeneration": (
"vllm.models.qwen4_exp",
"Qwen4ExpForConditionalGeneration",
),
"RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
"SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
"SmolVLMForConditionalGeneration": ("smolvlm", "SmolVLMForConditionalGeneration"),
@@ -659,6 +667,7 @@ _SPECULATIVE_DECODING_MODELS = {
"MedusaModel": ("medusa", "Medusa"),
"OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
"Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
"Qwen4ExpMTP": ("vllm.models.qwen4_exp", "Qwen4ExpMTP"),
"Step3p5MTP": ("step3p5_mtp", "Step3p5MTP"),
"Qwen3_5MTP": ("qwen3_5_mtp", "Qwen3_5MTP"),
"Qwen3_5MoeMTP": ("qwen3_5_mtp", "Qwen3_5MoeMTP"),