[Hardware] Support platforms and plugin system (#774)

Signed-off-by: gcanlin <canlinguosdu@gmail.com>
This commit is contained in:
Canlin Guo authored and GitHub committed 2026-01-28 22:12:52 +08:00
1 parent b11d436e19
commit c9392dff01
103 files changed
+1770 -1218

No files matched your search

@@ -36,7 +36,7 @@ from vllm_omni.model_executor.model_loader.weight_utils import download_weights_
from vllm_omni.model_executor.models.output_templates import OmniOutput
from vllm_omni.model_executor.models.utils import add_prefix_to_loaded_weights, split_list_into_ranges
from vllm_omni.model_executor.models.vision import get_llm_pos_ids_for_vision
from vllm_omni.utils.platform_utils import is_npu
from vllm_omni.platforms import current_omni_platform
TALKER_CODEC_EOS_TOKEN_ID = 8294
TALKER_CODEC_BOS_TOKEN_ID = 8293
@@ -247,7 +247,7 @@ class Qwen2_5OmniForConditionalGeneration(
if inputs_embeds is not None and inputs_embeds.device != thinker_dev:
inputs_embeds = inputs_embeds.to(thinker_dev)
if is_npu():
if current_omni_platform.is_npu():
# TODO: remove this hack when NPU supports batched inputs properly
thinker_input_ids = input_ids[0] if input_ids is not None and added_batch_dim else input_ids
# For MRoPE, positions shape is [3, num_tokens] (T/H/W), don't slice it
@@ -35,7 +35,7 @@ from vllm.v1.sample.metadata import SamplingMetadata
from vllm.v1.sample.sampler import Sampler
from vllm_omni.model_executor.models.qwen2_5_omni.audio_length import cap_and_align_mel_length, resolve_max_mel_frames
from vllm_omni.utils.platform_utils import is_npu
from vllm_omni.platforms import current_omni_platform
# Provide a no-op auto_docstring decorator to satisfy annotations if missing
@@ -732,7 +732,7 @@ def kaiser_sinc_filter1d(cutoff: float, half_width: float, kernel_size: int) ->
beta = 0.0
# TODO: When torch.kaiser_window supports NPU, remove the device="cpu" argument
if is_npu():
if current_omni_platform.is_npu():
kaiser_window = torch.kaiser_window(
kernel_size, beta=beta, periodic=False, dtype=torch.float32, device="cpu"
).to("npu")
@@ -796,7 +796,7 @@ class UpSample1d(nn.Module):
def forward(self, hidden_states):
channels = hidden_states.shape[1]
if is_npu():
if current_omni_platform.is_npu():
# TODO: When F.pad supports replicate mode on NPU, remove this branch
input_dtype = hidden_states.dtype
# F.pad in NPU doesn't support BF16 when mode is replicate.
@@ -843,7 +843,7 @@ class DownSample1d(nn.Module):
def forward(self, hidden_states):
channels = hidden_states.shape[1]
if is_npu():
if current_omni_platform.is_npu():
input_dtype = hidden_states.dtype
# F.pad in NPU doesn't support BF16 when mode is replicate.
# To ensure the accuracy, manually pad the input tensor.
@@ -35,7 +35,7 @@ from vllm.v1.sample.sampler import Sampler
from vllm_omni.model_executor.custom_process_mixin import CustomProcessMixin
from vllm_omni.model_executor.models.output_templates import OmniOutput
from vllm_omni.model_executor.models.utils import add_prefix_to_loaded_weights, safe_tensor_reshape
from vllm_omni.utils.platform_utils import is_npu
from vllm_omni.platforms import current_omni_platform
# Special token IDs for Qwen3 Omni MoE
# Reference: https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct/blob/main/tokenizer_config.json
@@ -269,7 +269,7 @@ class Qwen3OmniMoeForConditionalGeneration(
# ========== Stage 1: Thinker ==========
if self.model_stage == "thinker":
thinker_dev = self._module_device(self.thinker)
if is_npu():
if current_omni_platform.is_npu():
# Normalize to batched inputs if needed
_added_batch_dim = False
if input_ids is not None and input_ids.ndim == 1:
@@ -308,7 +308,7 @@ class Qwen3OmniMoeForConditionalGeneration(
"capture_layer_indices": [0, int(accept_layer)],
"return_hidden_states": True,
}
if is_npu():
if current_omni_platform.is_npu():
# TODO: remove this hack when NPU supports batched inputs properly
thinker_input_ids = input_ids[0] if input_ids is not None and _added_batch_dim else input_ids
thinker_inputs_embeds = (