[Bugfix] Fix DeepSeek V4 MTP HC state handling (#42320)

Signed-off-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>
This commit is contained in:
Mohammad Miadh Angkad
2026-05-13 15:44:52 -07:00
committed by GitHub
parent 597ed13803
commit f1cc7aad3c
2 changed files with 8 additions and 5 deletions
+4 -4
View File
@@ -1212,10 +1212,10 @@ class DeepseekV4DecoderLayer(nn.Module):
x: torch.Tensor,
positions: torch.Tensor,
input_ids: torch.Tensor | None,
post_mix: torch.Tensor | None,
res_mix: torch.Tensor | None,
residual: torch.Tensor | None,
) -> torch.Tensor:
post_mix: torch.Tensor | None = None,
res_mix: torch.Tensor | None = None,
residual: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
if residual is None:
# Run standalone hc_pre on first layer
residual = x
@@ -143,9 +143,12 @@ class DeepSeekV4MultiTokenPredictorLayer(nn.Module):
hidden_states = self.h_proj(previous_hidden_states) + self.e_proj(
inputs_embeds
).unsqueeze(-2)
hidden_states = self.mtp_block(
hidden_states, residual, post_mix, res_mix = self.mtp_block(
positions=positions, x=hidden_states, input_ids=None
)
hidden_states = self.mtp_block.hc_post(
hidden_states, residual, post_mix, res_mix
)
# Return the flat pre-hc_head residual so it can be re-fed as the
# next spec step's `previous_hidden_states` when
# num_speculative_tokens > 1. hc_head is deferred to compute_logits.