[Qwen4Exp] support Sequence Parallelism for Qwen3.8-Flash-Next (#56322)

Signed-off-by: huanghaoyan.hhy <huanghaoyan.hhy@alibaba-inc.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Harry Huang
2026-10-12 20:56:44 +08:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 258b97ff6b
commit c75cb63cd2
23 changed files with 870 additions and 178 deletions
@@ -27,6 +27,7 @@ from vllm.model_executor.layers.quantization.base_config import (
method_has_implemented_embedding,
resolve_quant_method,
)
from vllm.model_executor.layers.quantization.utils.fp8_utils import is_fp8
from vllm.model_executor.layers.utils import dispatch_unquantized_gemm
from vllm.model_executor.parameter import BasevLLMParameter
from vllm.model_executor.utils import set_weight_attrs
@@ -240,6 +241,7 @@ class VocabParallelEmbedding(PluggableLayer):
quant_config: quant config for the layer
prefix: full name of the layer in the state dict
disable_tp: If true, tensor parallelism will be disabled for this layer.
reduce_results: Whether to all-reduce contributions from vocabulary shards.
quant_method: Preselected quantization method for model-specific layers.
parallel_group: Process group used to shard and reduce the embedding.
@@ -258,6 +260,7 @@ class VocabParallelEmbedding(PluggableLayer):
prefix: str = "",
*,
disable_tp: bool = False,
reduce_results: bool = True,
quant_method: QuantizeMethodBase | None = None,
parallel_group: GroupCoordinator | None = None,
):
@@ -266,6 +269,7 @@ class VocabParallelEmbedding(PluggableLayer):
# Keep the input dimensions.
self.disable_tp = disable_tp
self.parallel_group = parallel_group
self.reduce_results = reduce_results
if disable_tp:
tp_rank, self.tp_size = 0, 1
elif parallel_group is not None:
@@ -516,6 +520,7 @@ class VocabParallelEmbedding(PluggableLayer):
param[loaded_weight.shape[0] :].data.fill_(0)
def forward(self, input_):
"""Look up IDs, optionally reducing contributions from vocabulary shards."""
if self.tp_size == 1:
return self.quant_method.embedding(self, input_.long())
@@ -543,24 +548,20 @@ class VocabParallelEmbedding(PluggableLayer):
self.shard_indices.added_vocab_end_index,
)
output_parallel = self.quant_method.embedding(self, masked_input.long())
if output_parallel.dtype in (
torch.float8_e4m3fn,
torch.float8_e5m2,
):
# Each vocab token has one owner, so FP8 bytes can use int8 SUM.
comm_output = output_parallel.view(torch.int8)
comm_output.masked_fill_(input_mask.unsqueeze(-1), 0)
output = (
self.parallel_group.all_reduce(comm_output)
if self.parallel_group is not None
else tensor_model_parallel_all_reduce(comm_output)
)
return output.view(output_parallel.dtype)
output_dtype = output_parallel.dtype
if is_fp8(output_dtype):
# Each vocab token has one owner, so FP8 bytes can use int8 SUM.
output_parallel = output_parallel.view(torch.int8)
if not self.use_fused_embedding:
output_parallel.masked_fill_(input_mask.unsqueeze(-1), 0)
# Reduce across all the model parallel GPUs.
if self.parallel_group is not None:
return self.parallel_group.all_reduce(output_parallel)
return tensor_model_parallel_all_reduce(output_parallel)
if self.reduce_results:
output_parallel = (
self.parallel_group.all_reduce(output_parallel)
if self.parallel_group is not None
else tensor_model_parallel_all_reduce(output_parallel)
)
return output_parallel.view(output_dtype)
def extra_repr(self) -> str:
s = f"num_embeddings={self.num_embeddings}"
+4
View File
@@ -3,6 +3,7 @@
from .interfaces import (
HasInnerState,
SupportsHCSP,
SupportsLoRA,
SupportsMRoPE,
SupportsMultiModal,
@@ -10,6 +11,7 @@ from .interfaces import (
SupportsPP,
SupportsTranscription,
has_inner_state,
supports_hc_sp,
supports_lora,
supports_mrope,
supports_multimodal,
@@ -33,6 +35,8 @@ __all__ = [
"is_text_generation_model",
"HasInnerState",
"has_inner_state",
"SupportsHCSP",
"supports_hc_sp",
"SupportsLoRA",
"supports_lora",
"SupportsMultiModal",
+10
View File
@@ -912,11 +912,21 @@ class Qwen4ExpForConditionalGenerationConfig(Qwen3_5ForConditionalGenerationConf
@staticmethod
def verify_and_update_config(vllm_config: "VllmConfig") -> None:
from vllm.platforms import current_platform
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
# MoE SP also enables HC SP without the explicit HC flag.
if (
current_platform.is_cuda()
and parallel_config.use_sequence_parallel_moe
and parallel_config.pipeline_parallel_size != 1
):
raise ValueError("Qwen4Exp MoE SP requires PP=1")
uses_ple_or_qsa = bool(text_config.ple_layer_ids) or (
getattr(text_config, "indexer_n_heads", None) is not None
)
+21
View File
@@ -858,6 +858,27 @@ def _supports_pp_inspect(model: type[object] | object) -> bool:
return supports_kw(model_forward, "intermediate_tensors")
@runtime_checkable
class SupportsHCSP(Protocol):
"""Support sequence-parallel HC without changing MoE parallelism."""
supports_hc_sp: ClassVar[Literal[True]] = True
@overload
def supports_hc_sp(model: type[object]) -> TypeIs[type[SupportsHCSP]]: ...
@overload
def supports_hc_sp(model: object) -> TypeIs[SupportsHCSP]: ...
def supports_hc_sp(
model: type[object] | object,
) -> TypeIs[type[SupportsHCSP]] | TypeIs[SupportsHCSP]:
return getattr(model, "supports_hc_sp", False)
@runtime_checkable
class HasInnerState(Protocol):
"""The interface required for all models that has inner state."""
+7 -1
View File
@@ -129,7 +129,12 @@ def _should_replicate_misaligned_shared_expert(
class Qwen3NextSparseMoeBlock(nn.Module):
def __init__(self, vllm_config: VllmConfig, prefix: str = ""):
def __init__(
self,
vllm_config: VllmConfig,
prefix: str = "",
reduce_results: bool = True,
):
super().__init__()
config = vllm_config.model_config.hf_text_config
@@ -231,6 +236,7 @@ class Qwen3NextSparseMoeBlock(nn.Module):
num_redundant_experts=self.n_redundant_experts,
is_sequence_parallel=self.is_sequence_parallel,
n_shared_experts=1 if self.shared_expert is None else None,
reduce_results=reduce_results,
fuse_shared_experts=self.is_fused_shared_expert_enabled,
shared_expert_prefix=f"{prefix}.shared_expert",
shared_expert_gate=self.shared_expert_gate
+12
View File
@@ -49,6 +49,7 @@ from .interfaces import (
is_attention_free,
is_hybrid,
requires_raw_input_tokens,
supports_hc_sp,
supports_mamba_prefix_caching,
supports_multimodal,
supports_multimodal_encoder_tp_data,
@@ -873,6 +874,7 @@ class _ModelInfo:
requires_raw_input_tokens: bool
supports_multimodal_encoder_tp_data: bool
supports_pp: bool
supports_hc_sp: bool
has_inner_state: bool
is_attention_free: bool
is_hybrid: bool
@@ -903,6 +905,7 @@ class _ModelInfo:
model
),
supports_pp=supports_pp(model),
supports_hc_sp=supports_hc_sp(model),
has_inner_state=has_inner_state(model),
is_attention_free=is_attention_free(model),
is_hybrid=is_hybrid(model),
@@ -1463,6 +1466,15 @@ class _ModelRegistry:
model_info, _ = self.inspect_model_cls(architectures, model_config)
return model_info.supports_pp
def is_hc_sp_supported_model(
self,
architectures: str | list[str],
model_config: ModelConfig,
) -> bool:
"""Check support for HC SP without changing MoE parallelism."""
model_info, _ = self.inspect_model_cls(architectures, model_config)
return model_info.supports_hc_sp
def model_has_inner_state(
self,
architectures: str | list[str],