[Bugfix] Fix DSV4 MTP after ROCm mHC integration (#42930)

Signed-off-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>
This commit is contained in:
Mohammad Miadh Angkad
2026-05-18 17:02:01 +00:00
committed by GitHub
parent 8c296de63b
commit 67f58ce23f
2 changed files with 17 additions and 12 deletions
+12 -8
View File
@@ -1261,10 +1261,12 @@ 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 | None, torch.Tensor | None, torch.Tensor | None
]:
residual = x
x, post, comb = self.hc_pre(
x, self.hc_attn_fn, self.hc_attn_scale, self.hc_attn_base
@@ -1288,10 +1290,12 @@ 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 | None, torch.Tensor | None, torch.Tensor | None
]:
if current_platform.is_rocm():
return self._forward_rocm(
x, positions, input_ids, post_mix, res_mix, residual
@@ -146,9 +146,10 @@ class DeepSeekV4MultiTokenPredictorLayer(nn.Module):
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
)
if current_platform.is_cuda():
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.
@@ -235,7 +236,7 @@ class DeepSeekV4MultiTokenPredictor(nn.Module):
hidden_states = hidden_states.view(
-1, mtp_layer.hc_mult, mtp_layer.config.hidden_size
)
hidden_states = self.hc_head_op(
hidden_states = mtp_layer.hc_head_op(
hidden_states,
mtp_layer.hc_head_fn,
mtp_layer.hc_head_scale,