biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[3/N] Initialize MM components in context managers (I-L) (#32650)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-01-20 18:21:56 +08:00
committed by GitHub
parent 8be263c3fb
commit 7f1bcd18ff
10 changed files with 203 additions and 389 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
vllm/model_executor/models/isaac.py
View File

@@ -1342,11 +1342,14 @@ class IsaacForConditionalGeneration(
"mrope_interleaved", rope_scaling["mrope_interleaved"]
)
target_cfg.rope_parameters = rope_parameters
self.language_model = init_vllm_registered_model(
vllm_config=vllm_config,
architectures=["Qwen3ForCausalLM"],
prefix=maybe_prefix(prefix, "language_model"),
)
with self._mark_language_model(vllm_config):
self.language_model = init_vllm_registered_model(
vllm_config=vllm_config,
architectures=["Qwen3ForCausalLM"],
prefix=maybe_prefix(prefix, "language_model"),
)
self.make_empty_intermediate_tensors = (
self.language_model.make_empty_intermediate_tensors
)
@@ -1363,14 +1366,16 @@ class IsaacForConditionalGeneration(
vision_cfg._attn_implementation = attn_impl
hidden_dim = vision_cfg.hidden_size * (vision_cfg.pixel_shuffle_scale_factor**2)
self.vision_embedding = IsaacVisionEmbedding(
vision_cfg=vision_cfg,
hidden_dim=hidden_dim,
output_dim=config.hidden_size,
quant_config=quant_config,
multimodal_config=self.multimodal_config,
prefix=maybe_prefix(prefix, "vision_embedding"),
)
with self._mark_tower_model(vllm_config, "image"):
self.vision_embedding = IsaacVisionEmbedding(
vision_cfg=vision_cfg,
hidden_dim=hidden_dim,
output_dim=config.hidden_size,
quant_config=quant_config,
multimodal_config=self.multimodal_config,
prefix=maybe_prefix(prefix, "vision_embedding"),
)
def iter_mm_grid_hw(
self, input_tokens: list[int], mm_features: list[MultiModalFeatureSpec]
@@ -1457,18 +1462,6 @@ class IsaacForConditionalGeneration(
return ()
return self._process_image_input(image_input)
def get_multimodal_embeddings(
self, **kwargs: object
) -> MultiModalEmbeddings | None:
# Backward compatibility for older runners.
embeddings = self.embed_multimodal(**kwargs)
if not embeddings:
return []
return embeddings
def get_language_model(self) -> torch.nn.Module:
return self.language_model
def forward(
self,
input_ids: torch.Tensor,