EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

move contiguous in fused_sigmoid_gating_delta_rule_update to model_runner_v1 (#5274)

Browse Source 
### What this PR does / why we need it?
The contiguous() operation temporarily increases memory usage, leading
to higher peak GPU memory, which necessitates reducing
gpu_memory_utilization. However, making tensors contiguous in
modelrunnerv1 significantly enhances operator performance, resulting in
greater end-to-end model benefits despite the memory overhead.

- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c

Signed-off-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
Co-authored-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
This commit is contained in:
XiaoxinWang
2025-12-26 09:19:47 +08:00
committed by GitHub
parent 9b2a7d8866
commit 320877d488
2 changed files with 15 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm_ascend/ops/triton/fla/sigmoid_gating.py
View File

@@ -360,7 +360,7 @@ def fused_sigmoid_gating_delta_rule_update(
if not initial_state_indices.is_contiguous(): if not initial_state_indices.is_contiguous():
initial_state_indices = initial_state_indices.contiguous() initial_state_indices = initial_state_indices.contiguous()
if not initial_state_source.is_contiguous(): if not initial_state_source.is_contiguous():
initial_state_source_contiguous = initial_state_source.contiguous() initial_state_source = initial_state_source.contiguous()
if not cu_seqlens.is_contiguous(): if not cu_seqlens.is_contiguous():
cu_seqlens = cu_seqlens.contiguous() cu_seqlens = cu_seqlens.contiguous()
@@ -375,7 +375,7 @@ def fused_sigmoid_gating_delta_rule_update(
v=v, v=v,
b=b, b=b,
o=o, o=o,
h0_source=initial_state_source_contiguous, h0_source=initial_state_source,
h0_indices=initial_state_indices, h0_indices=initial_state_indices,
cu_seqlens=cu_seqlens, cu_seqlens=cu_seqlens,
scale=scale, scale=scale,
@@ -391,7 +391,5 @@ def fused_sigmoid_gating_delta_rule_update(
num_warps=num_warps, num_warps=num_warps,
num_stages=num_stages, num_stages=num_stages,
) )
initial_state_source.copy_(
initial_state_source_contiguous.view_as(initial_state_source))
o = o.squeeze(0) o = o.squeeze(0)
return o return o

31
vllm_ascend/worker/model_runner_v1.py
View File

@@ -57,7 +57,6 @@ from vllm.sequence import IntermediateTensors
from vllm.utils.import_utils import LazyLoader from vllm.utils.import_utils import LazyLoader
from vllm.utils.math_utils import cdiv from vllm.utils.math_utils import cdiv
from vllm.utils.mem_utils import DeviceMemoryProfiler from vllm.utils.mem_utils import DeviceMemoryProfiler
from vllm.utils.torch_utils import get_dtype_size
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
from vllm.v1.attention.backends.utils import (AttentionCGSupport, from vllm.v1.attention.backends.utils import (AttentionCGSupport,
CommonAttentionMetadata) CommonAttentionMetadata)
@@ -2541,6 +2540,7 @@ class NPUModelRunner(GPUModelRunner):
) % kv_cache_spec.page_size_bytes == 0 ) % kv_cache_spec.page_size_bytes == 0
num_blocks = raw_tensor.numel( num_blocks = raw_tensor.numel(
) // kv_cache_spec.page_size_bytes ) // kv_cache_spec.page_size_bytes
assert num_blocks >= kv_cache_config.num_blocks
# `num_blocks` is the number of blocks the model runner can use. # `num_blocks` is the number of blocks the model runner can use.
# `kv_cache_config.num_blocks` is the number of blocks that # `kv_cache_config.num_blocks` is the number of blocks that
@@ -2549,27 +2549,22 @@ class NPUModelRunner(GPUModelRunner):
# different memory capacities, `num_blocks` can be different on # different memory capacities, `num_blocks` can be different on
# different GPUs, and `kv_cache_config.num_blocks` is set to # different GPUs, and `kv_cache_config.num_blocks` is set to
# the min of all `num_blocks`. Verify it here. # the min of all `num_blocks`. Verify it here.
assert num_blocks >= kv_cache_config.num_blocks
state_tensors = [] state_tensors = []
storage_offset_bytes = 0 target_idx = 0
for (shape, dtype) in zip(kv_cache_spec.shapes, start_idx = 0
kv_cache_spec.dtypes): for shape, dtype in zip(kv_cache_spec.shapes,
dtype_size = get_dtype_size(dtype) kv_cache_spec.dtypes):
num_element_per_page = ( # normally, there is conv state and ssm state in this loop. And there is only
kv_cache_spec.page_size_bytes // dtype_size) # a conv state in some special models.
target_shape = (num_blocks, *shape) target_shape = (num_blocks, *shape)
stride = torch.empty(target_shape).stride()
target_stride = (num_element_per_page, *stride[1:]) target_idx += torch.prod(
assert storage_offset_bytes % dtype_size == 0 torch.tensor(target_shape)).item()
tensor = torch.as_strided( tensor = raw_tensor.view(
raw_tensor.view(dtype), dtype)[start_idx:target_idx].view(target_shape)
size=target_shape, start_idx = target_idx
stride=target_stride,
storage_offset=storage_offset_bytes // dtype_size,
)
state_tensors.append(tensor) state_tensors.append(tensor)
storage_offset_bytes += stride[0] * dtype_size
kv_caches[layer_name] = state_tensors kv_caches[layer_name] = state_tensors
else: else:
raise ValueError("Unknown KV cache spec type.") raise ValueError("Unknown KV cache spec type.")