From cba06764d7a9da41e6f535d6355c13f725574f07 Mon Sep 17 00:00:00 2001 From: wangxiyuan Date: Wed, 19 Aug 2026 18:31:19 +0800 Subject: [PATCH] [Platform] Fill in the missing backend parameter for torch.compile (#51781) Signed-off-by: wangxiyuan --- vllm/model_executor/models/diffusion_gemma.py | 6 +++--- vllm/model_executor/models/kimi_k25_vit.py | 6 +++++- vllm/model_executor/models/parakeet.py | 7 ++++--- vllm/transformers_utils/processors/nano_nemotron_vl.py | 3 ++- vllm/v1/sample/ops/topk_topp_sampler.py | 2 +- 5 files changed, 15 insertions(+), 9 deletions(-) diff --git a/vllm/model_executor/models/diffusion_gemma.py b/vllm/model_executor/models/diffusion_gemma.py index 52a3824860..3a51e5aad6 100644 --- a/vllm/model_executor/models/diffusion_gemma.py +++ b/vllm/model_executor/models/diffusion_gemma.py @@ -126,7 +126,7 @@ class DiffusionGemmaProcessingInfo(Gemma4ProcessingInfo): return super().get_mm_max_tokens_per_item(seq_len, mm_counts) -@torch.compile(dynamic=True) +@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend) def _softcap_logits(logits: torch.Tensor, cap: float) -> torch.Tensor: # fp32 before tanh for numerical stability (matches HF DiffusionGemma). # Compiling fuses the cast/div/tanh/mul into one elementwise kernel over @@ -454,7 +454,7 @@ class DiffusionGemmaForConditionalGeneration( raise ValueError(f"Unsupported modality: {modality}") -@torch.compile(dynamic=True) +@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend) def _compute_num_rejected( num_logits: torch.Tensor, num_sampled: torch.Tensor, @@ -466,7 +466,7 @@ def _compute_num_rejected( return torch.where(is_denoise, query_lens, num_rejected) -@torch.compile(dynamic=True) +@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend) def _compiled_sample_step( # Logits from the model [num_decode * CL, vocab] logits: torch.Tensor, diff --git a/vllm/model_executor/models/kimi_k25_vit.py b/vllm/model_executor/models/kimi_k25_vit.py index dca62d596b..20f5b105b9 100644 --- a/vllm/model_executor/models/kimi_k25_vit.py +++ b/vllm/model_executor/models/kimi_k25_vit.py @@ -61,7 +61,11 @@ def get_rope_shape_decorate(func): @get_rope_shape_decorate -@torch.compile(dynamic=True, disable=current_platform.simple_compile_backend == "tpu") +@torch.compile( + dynamic=True, + backend=current_platform.simple_compile_backend, + disable=current_platform.simple_compile_backend == "tpu", +) def get_rope_shape(org, interpolation_mode, shape): return ( F.interpolate( diff --git a/vllm/model_executor/models/parakeet.py b/vllm/model_executor/models/parakeet.py index 5671ba05c5..ba7b96b824 100644 --- a/vllm/model_executor/models/parakeet.py +++ b/vllm/model_executor/models/parakeet.py @@ -19,6 +19,7 @@ from vllm.logger import init_logger from vllm.model_executor.layers.activation import ReLUSquaredActivation from vllm.model_executor.layers.layernorm import RMSNorm from vllm.model_executor.model_loader.weight_utils import default_weight_loader +from vllm.platforms import current_platform from vllm.transformers_utils.configs.parakeet import ExtractorConfig, ParakeetConfig logger = init_logger(__name__) @@ -186,7 +187,7 @@ class ParakeetExtractor: ) return self._apply_mel_filters(stft, mel_filters) - @torch.compile(dynamic=True) + @torch.compile(dynamic=True, backend=current_platform.simple_compile_backend) def _apply_mel_filters( self, stft_output: torch.Tensor, mel_filters: torch.Tensor ) -> torch.Tensor: @@ -195,7 +196,7 @@ class ParakeetExtractor: mel_spec = torch.log(mel_spec + LOG_ZERO_GUARD_VALUE) return mel_spec.permute(0, 2, 1) - @torch.compile(dynamic=True) + @torch.compile(dynamic=True, backend=current_platform.simple_compile_backend) def _apply_preemphasis( self, input_features: torch.Tensor, audio_lengths: torch.Tensor ) -> torch.Tensor: @@ -213,7 +214,7 @@ class ParakeetExtractor: input_features = input_features.masked_fill(~timemask, 0.0) return input_features - @torch.compile(dynamic=True) + @torch.compile(dynamic=True, backend=current_platform.simple_compile_backend) def _normalize_mel_features( self, mel_features: torch.Tensor, audio_lengths: torch.Tensor ) -> tuple[torch.Tensor, torch.Tensor]: diff --git a/vllm/transformers_utils/processors/nano_nemotron_vl.py b/vllm/transformers_utils/processors/nano_nemotron_vl.py index 028d207a25..022eb28f3f 100644 --- a/vllm/transformers_utils/processors/nano_nemotron_vl.py +++ b/vllm/transformers_utils/processors/nano_nemotron_vl.py @@ -26,6 +26,7 @@ from vllm.model_executor.models.parakeet import ParakeetExtractor from vllm.multimodal.inputs import AudioItem from vllm.multimodal.processing.processor import PromptUpdateDetails from vllm.multimodal.video_prune.evs import compute_retained_tokens_count +from vllm.platforms import current_platform from vllm.tokenizers.hf import HfTokenizer from .internvl import calculate_internvl_targets, get_internvl_target_ratios @@ -56,7 +57,7 @@ def calculate_timestamps( return timestamps -@torch.compile(dynamic=True) +@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend) def _bicubic_resize_and_normalize( tensor: torch.Tensor, size: tuple[int, int] | None = None, diff --git a/vllm/v1/sample/ops/topk_topp_sampler.py b/vllm/v1/sample/ops/topk_topp_sampler.py index c83ac7c4b9..e6e0e054c0 100644 --- a/vllm/v1/sample/ops/topk_topp_sampler.py +++ b/vllm/v1/sample/ops/topk_topp_sampler.py @@ -338,7 +338,7 @@ class TopKTopPSampler(nn.Module): # Note: this is a workaround for # https://github.com/pytorch/pytorch/pull/151218 -@torch.compile(dynamic=True) +@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend) def compiled_random_sample(logits: torch.Tensor) -> torch.Tensor: probs = logits.softmax(dim=-1, dtype=torch.float32) q = torch.empty_like(probs)