[Bugfix][MoE] Fix fused block-scale orientation (#50727)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Andreas Karatzas
2026-08-11 16:05:38 -04:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent fd04edef3e
commit 36f4630d89
5 changed files with 80 additions and 29 deletions
+11 -2
View File
@@ -132,6 +132,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
self,
vllm_config: VllmConfig,
prefix: str = "",
is_fused_checkpoint_transposed: bool = False,
):
super().__init__()
@@ -208,6 +209,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
enable_eplb=self.enable_eplb,
num_redundant_experts=self.n_redundant_experts,
is_sequence_parallel=self.is_sequence_parallel,
is_fused_checkpoint_transposed=is_fused_checkpoint_transposed,
)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
@@ -339,7 +341,12 @@ class Qwen3MoeAttention(nn.Module):
class Qwen3MoeDecoderLayer(nn.Module):
def __init__(self, vllm_config: VllmConfig, prefix: str = "") -> None:
def __init__(
self,
vllm_config: VllmConfig,
prefix: str = "",
is_fused_checkpoint_transposed: bool = False,
) -> None:
super().__init__()
config = vllm_config.model_config.hf_text_config
@@ -375,7 +382,9 @@ class Qwen3MoeDecoderLayer(nn.Module):
config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
):
self.mlp = Qwen3MoeSparseMoeBlock(
vllm_config=vllm_config, prefix=f"{prefix}.mlp"
vllm_config=vllm_config,
prefix=f"{prefix}.mlp",
is_fused_checkpoint_transposed=is_fused_checkpoint_transposed,
)
else:
self.mlp = Qwen3MoeMLP(
@@ -43,6 +43,7 @@ from vllm.tokenizers.registry import cached_tokenizer_from_config
from .interfaces import MixtureOfExperts
from .qwen3_moe import (
Qwen3MoeDecoderLayer,
Qwen3MoeForCausalLM,
Qwen3MoeModel,
Qwen3MoeSparseMoeBlock,
@@ -64,6 +65,15 @@ class Qwen3VLMoeProcessingInfo(Qwen3VLProcessingInfo):
return self.ctx.get_hf_config(Qwen3VLMoeConfig)
class Qwen3VLMoeDecoderLayer(Qwen3MoeDecoderLayer):
def __init__(self, vllm_config: VllmConfig, prefix: str = "") -> None:
super().__init__(
vllm_config=vllm_config,
prefix=prefix,
is_fused_checkpoint_transposed=True,
)
@support_torch_compile(
dynamic_arg_dims={
"input_ids": 0,
@@ -77,6 +87,19 @@ class Qwen3VLMoeProcessingInfo(Qwen3VLProcessingInfo):
}
)
class Qwen3MoeLLMModel(Qwen3MoeModel):
def __init__(
self,
*,
vllm_config: VllmConfig,
prefix: str = "",
decoder_layer_type: type[torch.nn.Module] = Qwen3VLMoeDecoderLayer,
):
super().__init__(
vllm_config=vllm_config,
prefix=prefix,
decoder_layer_type=decoder_layer_type,
)
def forward(
self,
input_ids: torch.Tensor | None,