xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-06-05 11:29:09 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Chore]: step3 forward_with_afd

Browse Source
This commit is contained in:
i-yuanyukun 2025-12-22 14:29:03 +08:00
parent 6d305dda38
commit 2a98ab3c8e
1 changed files with 60 additions and 31 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

91
vllm/model_executor/models/step3_text.py
View File

@ -398,43 +398,72 @@ class Step3TextModel(nn.Module):
positions: torch.Tensor, positions: torch.Tensor,
afd_metadata: AFDMetadata, afd_metadata: AFDMetadata,
) -> tuple[torch.Tensor, torch.Tensor]: ) -> tuple[torch.Tensor, torch.Tensor]:
forward_conext = get_forward_context()
recv_handle = None recv_handle = None
logger.info(f"{__file__}: forward with afd called, may blocked here")
for layer in islice(self.layers, self.start_layer, self.end_layer):
afd_connector = afd_metadata.afd_connector
afd_metadata.afd_stage_idx = dbo_current_ubatch_id()
if layer.layer_idx > 0: ubatch_hidden_states = []
hidden_states, recv_metadata = afd_connector.recv_ffn_output() ubatch_residual = []
if recv_metadata.recv_handle_list is not None:
recv_handle = recv_metadata.recv_handle_list
if recv_handle is not None: start_idx = 0
for work in recv_handle: for pos in afd_metadata.positions_list:
work.wait() num_tokens = pos.shape[1] if pos.ndim == 2 else pos.shape[0]
logger.info(f"Step3TextModel {layer.layer_idx=}: {hidden_states.shape=}, {positions.shape=}") end_idx = start_idx + num_tokens
current_hidden, residual = layer(positions, hidden_states, residual) ubatch_hidden_states.append(hidden_states[start_idx:end_idx])
logger.info(f"create attn metadata: {current_hidden.shape=}") ubatch_residual.append(
metadata = AFDConnectorMetadata.create_attention_metadata( residual[start_idx:end_idx] if residual is not None else None
layer_idx=layer.layer_idx,
stage_idx=afd_metadata.afd_stage_idx,
seq_len=current_hidden.shape[0],
dtype=current_hidden.dtype,
device=current_hidden.device,
num_of_stages=afd_metadata.num_of_stages,
afd_tokens_lens=afd_metadata.afd_tokens_lens,
) )
afd_connector.send_attn_output(current_hidden, metadata) start_idx = end_idx
if dbo_enabled(): for layer in islice(self.layers, self.start_layer, self.end_layer):
dbo_yield() for stage_i in range(forward_conext.afd_metadata.num_of_stages):
afd_connector = afd_metadata.afd_connector
forward_conext.attn_metadata = afd_metadata.attn_metadata_list[stage_i]
forward_conext.dp_metadata = afd_metadata.dp_metadata_list[stage_i]
hidden_states, recv_metadata = afd_connector.recv_ffn_output() residual = ubatch_residual[stage_i]
if recv_metadata.recv_handle_list is not None:
recv_handle = recv_metadata.recv_handle_list if layer.layer_idx > 0:
if recv_handle is not None: hidden_states, recv_metadata = afd_connector.recv_ffn_output()
for work in recv_handle: if recv_metadata.recv_handle_list is not None:
work.wait() recv_handle = recv_metadata.recv_handle_list
else:
hidden_states = ubatch_hidden_states[stage_i]
if recv_handle is not None:
for work in recv_handle:
work.wait()
current_positions = afd_metadata.positions_list[stage_i]
hidden_states, residual = layer(
current_positions, hidden_states, residual
)
ubatch_hidden_states[stage_i] = hidden_states
ubatch_residual[stage_i] = residual
logger.info(f"create attn metadata:, {afd_metadata.afd_tokens_lens=}")
metadata = AFDConnectorMetadata.create_attention_metadata(
layer_idx=layer.layer_idx,
stage_idx=stage_i,
seq_len=hidden_states.shape[0],
dtype=hidden_states.dtype,
device=hidden_states.device,
num_of_stages=afd_metadata.num_of_stages,
afd_tokens_lens=afd_metadata.afd_tokens_lens,
)
afd_connector.send_attn_output(hidden_states, metadata)
# Recv last layer FFN output.
for stage_i in range(afd_metadata.num_of_stages):
ubatch_hidden_states[stage_i], recv_metadata = (
afd_connector.recv_ffn_output()
)
# Re-assemble the batch
hidden_states = torch.cat(ubatch_hidden_states, dim=0)
if any(r is not None for r in ubatch_residual):
residual = torch.cat(ubatch_residual, dim=0)
else:
residual = None
return hidden_states, residual return hidden_states, residual