[Attention Backend] add HPC-Ops Attention backend (#46020)

Signed-off-by: chengvjiang <chengvjiang@tencent.com>
Co-authored-by: chengvjiang <chengvjiang@tencent.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Cheng Jiang
2026-06-30 18:17:43 +08:00
committed by GitHub
co-authored by chengvjiang Andreas Karatzas
parent 8e9d70fdd5
commit 8cf7c4d8ad
10 changed files with 978 additions and 14 deletions
+10
View File
@@ -20,6 +20,7 @@ from vllm.model_executor.layers.attention import (
MLAAttention,
MMEncoderAttention,
)
from vllm.model_executor.layers.hpc import HpcModule
from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig,
QuantizeMethodBase,
@@ -125,6 +126,15 @@ def process_weights_after_loading(
with device_loading_context(module, target_device):
module.process_weights_after_loading(model_config.dtype)
# Process HPC modules (HpcRopeNorm, etc.) that rely on
# process_weights_after_loading being called from the model's
# load_weights(). When using DummyModelLoader (e.g. profiling or
# sleep/wake_up reload), the model's load_weights() is not called, so we
# must handle HPC modules here generically.
for _, module in model.named_modules():
if isinstance(module, HpcModule):
module.process_weights_after_loading(model)
# Needed for torchao model reloading via model.reload_weights
# @kylesayrs @jerryzh168 this can be removed if callers move to `reload_weights`
if model_config.quantization == "torchao":