From e7a3963339654aa18b713f9ff233f129af9c7c8b Mon Sep 17 00:00:00 2001 From: Woosuk Kwon Date: Sun, 13 Sep 2026 02:23:20 -0700 Subject: [PATCH] [Bugfix] Avoid repeated dummy initialization and random CPU Engram fills (#56682) Signed-off-by: Woosuk Kwon --- vllm/model_executor/model_loader/dummy_loader.py | 8 ++++++-- vllm/model_executor/model_loader/weight_utils.py | 4 ++++ vllm/models/deepseek_v4_1/common/engram.py | 4 ++++ 3 files changed, 14 insertions(+), 2 deletions(-) diff --git a/vllm/model_executor/model_loader/dummy_loader.py b/vllm/model_executor/model_loader/dummy_loader.py index 98f82a5b71..0f263fe117 100644 --- a/vllm/model_executor/model_loader/dummy_loader.py +++ b/vllm/model_executor/model_loader/dummy_loader.py @@ -14,7 +14,6 @@ from vllm.model_executor.model_loader.reload.meta import materialize_layer from vllm.model_executor.model_loader.reload.types import LayerReloadingInfo from vllm.model_executor.model_loader.reload.utils import get_layer_tensors from vllm.model_executor.model_loader.weight_utils import ( - initialize_dummy_weights, initialize_single_dummy_weight, ) @@ -41,7 +40,12 @@ class DummyModelLoader(BaseModelLoader): else: # NOTE(woosuk): For accurate performance evaluation, we assign # random values to the weights. - initialize_dummy_weights(layer, model_config) + # Initialize only local tensors: state_dict() would recursively + # initialize descendants that model.modules() also visits. + # Exclude nonpersistent buffers, matching state_dict(). + for name, tensor in get_layer_tensors(layer).items(): + if name not in layer._non_persistent_buffers_set: + initialize_single_dummy_weight(tensor) def _process_online_quant_layer( self, diff --git a/vllm/model_executor/model_loader/weight_utils.py b/vllm/model_executor/model_loader/weight_utils.py index 3085df85c6..4cb8249e48 100644 --- a/vllm/model_executor/model_loader/weight_utils.py +++ b/vllm/model_executor/model_loader/weight_utils.py @@ -1341,6 +1341,10 @@ def initialize_single_dummy_weight( if param.device.type == "meta": return # deferred to finalize_layerwise_processing (e.g. online quant) + if (dummy_weight_value := getattr(param, "dummy_weight_value", None)) is not None: + param.fill_(dummy_weight_value) + return + if not torch.is_floating_point(param): if current_platform.is_rocm(): # On ROCm, integer params (e.g. GPTQ qweight/qzeros) are left diff --git a/vllm/models/deepseek_v4_1/common/engram.py b/vllm/models/deepseek_v4_1/common/engram.py index f30c3a5606..4cfba842b0 100644 --- a/vllm/models/deepseek_v4_1/common/engram.py +++ b/vllm/models/deepseek_v4_1/common/engram.py @@ -687,6 +687,10 @@ class ParallelEngramEmbedding(nn.Module): }, ) if cpu_offload: + # Constant dummy values avoid randomizing huge CPU lookup tables. + set_weight_attrs(self.weight, {"dummy_weight_value": 1.0}) + # The ue8m0 encoding of scale 1.0 is exponent byte 127. + set_weight_attrs(self.weight_scale_inv, {"dummy_weight_value": 127}) logger.info( "Engram table offloaded to pinned host memory: %d rows x %d, " "%.2f GiB per rank",