Remove unnecessary numpy imports and update return type in extract_multimodal_outputs method in OmniGPUModelRunner

- Eliminated redundant numpy import statements to streamline the code.
- Changed the return type of extract_multimodal_outputs from dict to tuple for better clarity on output structure.
This commit is contained in:
tzhouam committed 2025-10-24 11:08:42 +08:00
1 parent 4609afa613
commit ee4822327b
1 file changed
+1 -4
+1 -4
View File
@@ -123,7 +123,6 @@ class OmniGPUModelRunner(GPUModelRunner):
try:
if getattr(new_req_data, "prompt_embeds", None) is not None:
payload = new_req_data.prompt_embeds
import numpy as np
dtype = getattr(np, payload.dtype)
arr = np.frombuffer(payload.data, dtype=dtype)
arr = arr.reshape(payload.shape)
@@ -147,7 +146,6 @@ class OmniGPUModelRunner(GPUModelRunner):
else:
from vllm.v1.engine import AdditionalInformationPayload
if isinstance(payload_info, AdditionalInformationPayload):
import numpy as np
for k, entry in payload_info.entries.items():
if entry.tensor_data is not None:
dt = np.dtype(getattr(entry, "tensor_dtype", "float32"))
@@ -613,7 +611,7 @@ class OmniGPUModelRunner(GPUModelRunner):
)
@torch.inference_mode()
def extract_multimodal_outputs(self, hidden_states: torch.Tensor) -> dict:
def extract_multimodal_outputs(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, dict]:
if hasattr(self.model, "have_multimodal_outputs") and self.model.have_multimodal_outputs:
text_hidden_states = hidden_states.text_hidden_states
multimodal_outputs = hidden_states.multimodal_outputs
@@ -736,7 +734,6 @@ class OmniGPUModelRunner(GPUModelRunner):
else:
hidden_states = outputs
logger.warning(f"Multimodal outputs are not returned in the dummy run, need to double check the implementation!")
text_hidden_states, multimodal_outputs = self.extract_multimodal_outputs(hidden_states)