[NPU] upgrade to v0.30.0 (#8027)

Signed-off-by: Weiming Liao <liaowm5@gmail.com>
This commit is contained in:
Weiming Liao authored and GitHub committed 2026-09-23 14:59:29 +00:00
1 parent d319d276f9
commit 81bc7116c1
6 files changed
+22 -21

No files matched your search

@@ -544,7 +544,7 @@ class NPUARModelRunner(OmniNPUModelRunner, OmniConnectorModelRunnerMixin, Duplex
self.input_batch,
self.requests,
self.compilation_config.static_forward_context,
self.model.get_mamba_state_copy_func(),
self._get_mamba_state_copy_funcs(),
preprocess_bufs,
)
# preprocess_mamba resets num_accepted_tokens_cpu to 1
@@ -269,7 +269,7 @@ class NPUGenerationModelRunner(OmniNPUModelRunner, OmniConnectorModelRunnerMixin
self.input_batch,
self.requests,
self.compilation_config.static_forward_context,
self.model.get_mamba_state_copy_func(),
self._get_mamba_state_copy_funcs(),
preprocess_bufs,
)
# preprocess_mamba resets num_accepted_tokens_cpu to 1
@@ -854,8 +854,6 @@ class NPUGenerationModelRunner(OmniNPUModelRunner, OmniConnectorModelRunnerMixin
if self.uses_mrope:
positions = self.mrope_positions.gpu[:, :num_tokens_padded]
elif self.uses_xdrope_dim > 0:
positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
else:
positions = self.positions[:num_tokens_padded]
@@ -1,6 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
from contextlib import AbstractContextManager
from functools import partial
from typing import Any
@@ -35,9 +36,13 @@ else:
class OmniNPUModelRunner(OmniGPUModelRunner, NPUModelRunner):
def initialize_kv_cache(self, kv_cache_config) -> None:
def initialize_kv_cache(
self, kv_cache_config, kv_cache_allocation_context: AbstractContextManager | None = None
) -> None:
"""Stage the omni prefix-cache config (hidden / mm tensors reused on hits)."""
NPUModelRunner.initialize_kv_cache(self, kv_cache_config)
NPUModelRunner.initialize_kv_cache(
self, kv_cache_config, kv_cache_allocation_context=kv_cache_allocation_context
)
if getattr(self, "_omni_prefix_cache_cfg", None) is None:
# Same gate as the GPU runner (pooling stage, kv_consumer /
# kv_both, hybrid kv groups). Read the config back off
@@ -332,8 +337,6 @@ class OmniNPUModelRunner(OmniGPUModelRunner, NPUModelRunner):
if self.uses_mrope:
positions = self.mrope_positions.gpu[:, :num_tokens_padded]
elif self.uses_xdrope_dim > 0:
positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
else:
positions = self.positions[:num_tokens_padded]