[Bugfix][Model][Spec Decode] Defer disposable GLM MTP head (#55442)

Signed-off-by: Luca Motz <luca.motz@icloud.com>
This commit is contained in:
lucamotz
2026-09-22 20:08:27 +00:00
committed by GitHub
parent 71a2392845
commit 276db94fb2
3 changed files with 40 additions and 8 deletions
+11 -6
View File
@@ -54,15 +54,20 @@ class SharedHead(nn.Module):
config: PretrainedConfig,
prefix: str,
quant_config: QuantizationConfig | None = None,
*,
defer_lm_head: bool = False,
) -> None:
super().__init__()
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.head = ParallelLMHead(
config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=maybe_prefix(prefix, "head"),
)
if defer_lm_head:
self.head = None
else:
self.head = ParallelLMHead(
config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=maybe_prefix(prefix, "head"),
)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
return self.norm(hidden_states)