From 397094da1768c7a6f29dfa4f70079d793ac747df Mon Sep 17 00:00:00 2001 From: Lucain Date: Wed, 5 Aug 2026 15:49:56 +0200 Subject: [PATCH] Resolve revision to commit_hash once per model load, via huggingface_hub's `resolve_revision` (#49990) Signed-off-by: Wauplin Signed-off-by: Lucain Signed-off-by: Lucain Pouget Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: OpenAI Codex Co-authored-by: Claude Opus 5 (1M context) Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> --- requirements/common.txt | 1 + requirements/test/cpu.txt | 3 +- requirements/test/cuda.txt | 4 ++- requirements/test/rocm.txt | 4 ++- requirements/test/xpu.txt | 4 ++- vllm/config/model.py | 27 ++++++++++++++++++ vllm/model_executor/models/ultravox.py | 6 ++-- vllm/transformers_utils/repo_utils.py | 38 ++++++++++++++++++++++++++ vllm/v1/worker/gpu_model_runner.py | 3 ++ 9 files changed, 84 insertions(+), 6 deletions(-) diff --git a/requirements/common.txt b/requirements/common.txt index 748ffe88f4..2ccfc1960f 100644 --- a/requirements/common.txt +++ b/requirements/common.txt @@ -8,6 +8,7 @@ tqdm blake3 py-cpuinfo transformers >= 5.5.3 +huggingface_hub >= 1.26.0 tokenizers >= 0.21.1 # Required for fast incremental detokenization. safetensors >= 0.6.2 # MXFP4/MXFP6 dtype support (F8_E8M0, F4) added in 0.6.0: https://github.com/huggingface/safetensors/pull/611 protobuf >= 5.29.6, !=6.30.*, !=6.31.*, !=6.32.*, !=6.33.0.*, !=6.33.1.*, !=6.33.2.*, !=6.33.3.*, !=6.33.4.* # Required by LlamaTokenizer, gRPC. CVE-2026-0994 diff --git a/requirements/test/cpu.txt b/requirements/test/cpu.txt index aed7dd09a9..0237ba165b 100644 --- a/requirements/test/cpu.txt +++ b/requirements/test/cpu.txt @@ -341,8 +341,9 @@ httpx==0.27.2 # schemathesis httpx-sse==0.4.3 # via mcp -huggingface-hub==1.22.0 +huggingface-hub==1.26.0 # via + # -r requirements/test/../common.txt # accelerate # datasets # evaluate diff --git a/requirements/test/cuda.txt b/requirements/test/cuda.txt index 07d18e4f8a..4952fba64d 100644 --- a/requirements/test/cuda.txt +++ b/requirements/test/cuda.txt @@ -360,8 +360,10 @@ httpx==0.27.2 # schemathesis httpx-sse==0.4.3 # via mcp -huggingface-hub==1.22.0 +huggingface-hub==1.26.0 # via + # -c requirements/common.txt + # -r requirements/test/../common.txt # accelerate # datasets # evaluate diff --git a/requirements/test/rocm.txt b/requirements/test/rocm.txt index c7979e4efa..43283f8ff6 100644 --- a/requirements/test/rocm.txt +++ b/requirements/test/rocm.txt @@ -351,8 +351,10 @@ httpx==0.27.2 # schemathesis httpx-sse==0.4.3 # via mcp -huggingface-hub==1.22.0 +huggingface-hub==1.26.0 # via + # -c requirements/common.txt + # -r requirements/test/../common.txt # accelerate # datasets # evaluate diff --git a/requirements/test/xpu.txt b/requirements/test/xpu.txt index 4ecf0bede9..53410f4db5 100644 --- a/requirements/test/xpu.txt +++ b/requirements/test/xpu.txt @@ -224,8 +224,10 @@ httpx==0.28.1 # schemathesis httpx-sse==0.4.3 # via mcp -huggingface-hub==1.22.0 +huggingface-hub==1.26.0 # via + # -c requirements/common.txt + # -r requirements/test/../common.txt # accelerate # datasets # evaluate diff --git a/vllm/config/model.py b/vllm/config/model.py index d45f775217..517bb3847c 100644 --- a/vllm/config/model.py +++ b/vllm/config/model.py @@ -47,6 +47,7 @@ from vllm.transformers_utils.model_arch_config_convertor import ( MODEL_ARCH_CONFIG_CONVERTORS, ModelArchConfigConvertorBase, ) +from vllm.transformers_utils.repo_utils import resolve_revision from vllm.transformers_utils.runai_utils import ObjectStorageModel, is_runai_obj_uri from vllm.transformers_utils.utils import maybe_model_redirect from vllm.utils.import_utils import LazyLoader @@ -529,6 +530,7 @@ class ModelConfig: self.served_model_name = get_served_model_name( self.model, self.served_model_name ) + requested_revision = self.revision self.model = maybe_model_redirect(self.model) # The tokenizer is consistent with the model by default. if self.tokenizer is None: @@ -558,6 +560,31 @@ class ModelConfig: self.maybe_pull_model_tokenizer_for_runai(self.model, self.tokenizer) + # If loading model/tokenizer from HF Hub, resolve the revision once + # to prevent resolving it multiple times downstream. + can_resolve_model_revision = ( + self.hf_config_path is None or self.hf_config_path == self.model + ) + if can_resolve_model_revision: + self.revision = resolve_revision( + self.model, + self.revision, + self.hf_token, + ) + + if ( + can_resolve_model_revision + and self.tokenizer == self.model + and self.tokenizer_revision == requested_revision + ): + self.tokenizer_revision = self.revision + else: + self.tokenizer_revision = resolve_revision( + self.tokenizer, + self.tokenizer_revision, + self.hf_token, + ) + if self.override_attention_dtype is not None and not current_platform.is_rocm(): warnings.warn( "override-attention-dtype is set but not using ROCm platform", diff --git a/vllm/model_executor/models/ultravox.py b/vllm/model_executor/models/ultravox.py index 2c1b2b02e8..037b57bc1e 100644 --- a/vllm/model_executor/models/ultravox.py +++ b/vllm/model_executor/models/ultravox.py @@ -544,6 +544,8 @@ class UltravoxModel(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA): [self.config.audio_token_index], ) + # The towers live in their own Hub repos, so the revision of the + # Ultravox repo does not apply to them. self.secondary_weights = [] if config.audio_model_id is not None: # this prefix is not for initialization, but for loading weights @@ -551,7 +553,7 @@ class UltravoxModel(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA): self.secondary_weights.append( DefaultModelLoader.Source( model_or_path=config.audio_model_id, - revision=vllm_config.model_config.revision, + revision=None, prefix="audio_tower.", ) ) @@ -561,7 +563,7 @@ class UltravoxModel(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA): self.secondary_weights.append( DefaultModelLoader.Source( model_or_path=config.text_model_id, - revision=vllm_config.model_config.revision, + revision=None, prefix="language_model.", ) ) diff --git a/vllm/transformers_utils/repo_utils.py b/vllm/transformers_utils/repo_utils.py index a758f5d535..4066edbbf0 100644 --- a/vllm/transformers_utils/repo_utils.py +++ b/vllm/transformers_utils/repo_utils.py @@ -49,6 +49,44 @@ def hf_fs() -> "huggingface_hub.HfFileSystem": ) +def resolve_revision( + repo_id: str, + revision: str | None = None, + token: str | bool | None = None, +) -> str | None: + """Best-effort attempt to resolve a Hub revision to a commit hash. + + Resolving the commit hash once prevents repeated HTTP calls downstream and + avoids races if the branch moves while the model is loading. + + `HfApi.resolve_revision` returns a `ResolvedRevision`: a `str` equal to the + requested revision that also carries the commit hash, so downstream error + messages stay readable and offline loads reuse the cached `refs/` entry. + + Returns: + The resolved revision, or `revision` unchanged if it cannot be resolved + (local path, ModelScope, or any Hub error). + """ + if Path(repo_id).exists() or envs.VLLM_USE_MODELSCOPE: + return revision + + try: + return hf_api().resolve_revision( + repo_id, + revision=revision, + local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE, + token=token, + ) + except Exception: + logger.debug( + "Failed to resolve revision for %s; falling back to %s.", + repo_id, + revision, + exc_info=True, + ) + return revision + + _R = TypeVar("_R") diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py index d500b6720a..8585133bf0 100644 --- a/vllm/v1/worker/gpu_model_runner.py +++ b/vllm/v1/worker/gpu_model_runner.py @@ -5553,7 +5553,10 @@ class GPUModelRunner( ) if weights_path is not None: + # The revision belongs to the model we are reloading away from, + # so it must not be carried over to the new path. self.model_config.model = weights_path + self.model_config.revision = None weights_iterator = model_loader.get_all_weights(self.model_config, model) weights_iterator = cast( Iterable[tuple[str, torch.Tensor]], weights_iterator