[Quantization] Enable shared expert fusion compatibility with online shared_expert quantization (showcase: along Quark MXFP4 routed experts) (#55686)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
Signed-off-by: fxmarty-amd <felmarty@amd.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Misha Goin <mgoin64@gmail.com>
This commit is contained in:
fxmarty-amd
2026-10-08 12:04:46 -04:00
committed by GitHub
co-authored by OpenAI Codex Misha Goin
parent e2414ec1db
commit 9f3d70e081
24 changed files with 984 additions and 73 deletions
+1
View File
@@ -199,6 +199,7 @@ class AXK1MoE(nn.Module):
if self.is_fused_shared_expert_enabled
else None,
fuse_shared_experts=self.is_fused_shared_expert_enabled,
shared_expert_prefix=f"{prefix}.shared_experts",
)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
@@ -396,6 +396,7 @@ class DeepseekV2MoE(nn.Module):
if self.is_fused_shared_expert_enabled
else None,
fuse_shared_experts=self.is_fused_shared_expert_enabled,
shared_expert_prefix=f"{prefix}.shared_experts",
router_logits_dtype=self.gate.out_dtype,
)
+1
View File
@@ -204,6 +204,7 @@ class Glm4MoE(nn.Module):
config.n_shared_experts if self.is_fused_shared_expert_enabled else None
),
fuse_shared_experts=self.is_fused_shared_expert_enabled,
shared_expert_prefix=f"{prefix}.shared_experts",
)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+1
View File
@@ -232,6 +232,7 @@ class Qwen3NextSparseMoeBlock(nn.Module):
is_sequence_parallel=self.is_sequence_parallel,
n_shared_experts=1 if self.shared_expert is None else None,
fuse_shared_experts=self.is_fused_shared_expert_enabled,
shared_expert_prefix=f"{prefix}.shared_expert",
shared_expert_gate=self.shared_expert_gate
if self.shared_expert is None
else None,