Merge remote-tracking branch 'origin/main' into dev/vllm-align

This commit is contained in:
tzhouam committed 2026-02-20 01:38:44 +00:00
commit 663976b8ca
2 files changed
+29 -43

No files matched your search

@@ -1588,20 +1588,6 @@ class HunyuanFusedMoE(SharedFusedMoE):
self._init_hook_handle.remove() self._init_hook_handle.remove()
def forward(self, hidden_states, router_logits): def forward(self, hidden_states, router_logits):
from vllm.model_executor.layers.fused_moe.layer import get_forward_context
ctx = get_forward_context()
if not ctx.remaining_moe_layers:
import re
moe_names = [name for name in ctx.no_compile_layers.keys() if ".mlp.experts" in name]
def get_layer_num(name):
match = re.search(r"layers\.(\d+)\.mlp", name)
return int(match.group(1)) if match else -1
moe_names.sort(key=get_layer_num, reverse=True)
ctx.remaining_moe_layers.extend(moe_names)
return super().forward(hidden_states, router_logits) return super().forward(hidden_states, router_logits)
@@ -804,17 +804,14 @@ class HunyuanImage3Pipeline(HunyuanImage3PreTrainedModel, GenerationMixin):
) )
kwargs["num_image_tokens"] = num_image_tokens kwargs["num_image_tokens"] = num_image_tokens
# 50 and 5.0 hard code # 50 and 5.0 hard code
from vllm.forward_context import set_forward_context results = self.pipeline(
batch_size=len(batch_gen_image_info),
with set_forward_context(None, self.vllm_config): image_size=[batch_gen_image_info[0].image_height, batch_gen_image_info[0].image_width],
results = self.pipeline( num_inference_steps=kwargs.get("num_inference_steps", 50),
batch_size=len(batch_gen_image_info), guidance_scale=kwargs.get("guidance_scale", 5.0),
image_size=[batch_gen_image_info[0].image_height, batch_gen_image_info[0].image_width], generator=generator,
num_inference_steps=kwargs.get("num_inference_steps", 50), model_kwargs=kwargs,
guidance_scale=kwargs.get("guidance_scale", 5.0), )
generator=generator,
model_kwargs=kwargs,
)
samples = results[0] samples = results[0]
return samples return samples
@@ -929,24 +926,27 @@ class HunyuanImage3Pipeline(HunyuanImage3PreTrainedModel, GenerationMixin):
) )
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
outputs = self.model( from vllm.forward_context import set_forward_context
input_ids=input_ids,
attention_mask=attention_mask, with set_forward_context(None, self.vllm_config):
position_ids=position_ids, outputs = self.model(
past_key_values=past_key_values, input_ids=input_ids,
inputs_embeds=inputs_embeds, attention_mask=attention_mask,
use_cache=use_cache, position_ids=position_ids,
output_attentions=output_attentions, past_key_values=past_key_values,
output_hidden_states=output_hidden_states, inputs_embeds=inputs_embeds,
return_dict=return_dict, use_cache=use_cache,
custom_pos_emb=custom_pos_emb, output_attentions=output_attentions,
mode=mode, output_hidden_states=output_hidden_states,
first_step=first_step, return_dict=return_dict,
query_lens=query_lens, custom_pos_emb=custom_pos_emb,
seq_lens=seq_lens, mode=mode,
num_image_tokens=num_image_tokens, first_step=first_step,
gen_timestep_scatter_index=gen_timestep_scatter_index, query_lens=query_lens,
) seq_lens=seq_lens,
num_image_tokens=num_image_tokens,
gen_timestep_scatter_index=gen_timestep_scatter_index,
)
hidden_states = outputs[0] hidden_states = outputs[0]
if mode == "gen_text": if mode == "gen_text":