fix(step3p5-mtp): honor exclude_modules for the MTP head via prefix (#48883)

Signed-off-by: Chan Nguyen <channguyen@nvidia.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Chanh Nguyen
2026-07-29 16:58:29 +00:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 82553692b8
commit f98061ce6c
+11 -2
View File
@@ -32,11 +32,18 @@ class SharedHead(nn.Module):
self,
config: PretrainedConfig,
quant_config: QuantizationConfig | None = None,
prefix: str = "",
) -> None:
super().__init__()
self.norm = GemmaRMSNorm(config.hidden_size, config.rms_norm_eps)
# Give the head its prefix so the quant config's exclude_modules matcher
# can skip it; without one it defaults to "" and never matches, so a
# checkpoint-excluded (BF16) MTP head gets quantized -> load crash.
self.head = ParallelLMHead(
config.vocab_size, config.hidden_size, quant_config=quant_config
config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=f"{prefix}.head",
)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
@@ -55,7 +62,9 @@ class Step3p5AMultiTokenPredictorLayer(nn.Module):
self.enorm = GemmaRMSNorm(config.hidden_size, config.rms_norm_eps)
self.hnorm = GemmaRMSNorm(config.hidden_size, config.rms_norm_eps)
self.eh_proj = nn.Linear(config.hidden_size * 2, config.hidden_size, bias=False)
self.shared_head = SharedHead(config=config, quant_config=quant_config)
self.shared_head = SharedHead(
config=config, quant_config=quant_config, prefix=f"{prefix}.shared_head"
)
self.mtp_block = Step3p5DecoderLayer(
vllm_config,
prefix=f"{prefix}.mtp_block",