### What this PR does / why we need it?
The same fix from https://github.com/vllm-project/vllm/pull/36013.
In _update_states_after_model_execute, num_accepted_tokens is copied
from GPU to pinned CPU memory with non_blocking=True. The CPU-side numpy
view is later read in _build_attention_metadata during the next
execute_model call. With async scheduling, _bookkeeping_sync
deliberately avoids any CUDA synchronization (the whole point of async
scheduling), so there is no guarantee the DMA has landed before the CPU
read.
Signed-off-by: ppppeng <zepengliu912@qq.com>
### What this PR does / why we need it?
Fix TTFT degradation on Deepseek-V3.1-W4A8. Revert change of
`balance_flag` in https://github.com/vllm-project/vllm-ascend/pull/7611.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
- vLLM version: v0.18.0
Signed-off-by: Wangbingjie <wangbj1207@126.com>
### What this PR does / why we need it?
This PR introduces a caching mechanism for CPU-based `torch.Generator`
objects in the `_random_sample_310p` function to optimize sampling
performance. It includes unit tests for cache persistence and state
recovery. Feedback highlights a critical bug where keying the cache by
batch index instead of generator ID can break RNG reproducibility during
request re-scheduling, and notes a potential memory leak in the global
cache.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Tested via new unit tests in `tests/ut/_310p/sample/test_sampler_310.py`
verifying cache logic and error handling.
---------
Signed-off-by: csoulnd <daidaicurry@foxmail.com>
<!-- Thanks for sending a pull request!
BEFORE SUBMITTING, PLEASE READ
https://docs.vllm.ai/en/latest/contributing/overview.html
-->
### What this PR does / why we need it?
<!--
- Please clarify what changes you are proposing. The purpose of this
section is to outline the changes and how this PR fixes the issue.
If possible, please consider writing useful notes for better and faster
reviews in your PR.
- Please clarify why the changes are needed. For instance, the use case
and bug description.
- Fixes #
-->
The triton kernels in sample encounter some problems, scenarios are
shown below:
1. 【expand_kernel/ rejection_random_sample_kernel/
prepare_inputs_padded_kernel】, these three operations will use
‘tl.load(prt + offsets -1, mask)’ in their implementations, but triton
compiler reports that the masks in these scenarios are not static and
contiguous. As a result, compiler will first access this memory and
apply the mask. Therefore, I modified the code to ‘tl.load(prt
+tl.maximum(offsets - 1, 0), mask)’ to ensure no -1 reads.
2. 【sample_recovered_tokens_kernel/ rejection_random_sample_kernel】,
this kernel uses draft_token_id as an address offset for the load
operation. In the PD separation scenario, if the pad token is -1,
illegal memory reads and writes can occur. Therefore, i modified the
kernel and so they can do well with -1 token.
### Does this PR introduce _any_ user-facing change?
<!--
Note that it means *any* user-facing change including all aspects such
as API, interface or other behavior changes.
Documentation-only updates are not considered user-facing changes.
-->
### How was this patch tested?
<!--
CI passed with new added/existing test.
If it was tested in a way different from regular unit tests, please
clarify how you tested step by step, ideally copy and paste-able, so
that other reviewers can test and check, and descendants can verify in
the future.
If tests were not added, please describe why they were not added and/or
why it was difficult to add.
-->
Signed-off-by: ppppeng <zepengliu912@qq.com>
Co-authored-by: zepengliu912@qq.com <root@localhost.localdomain>
### What this PR does / why we need it?
Backport validate pd mode feature gates no fused mc2 v0.18.0 clean
backport #8582
---------
Signed-off-by: wangxiaoteng <wangxiaoteng@huawei.com>
### What this PR does / why we need it?
This backports the forced-tool-choice `content=None` guard to the
`releases/v0.18.0` compatibility layer.
Upstream vLLM still has forced named tool-choice branches that assert
`content is not None` after reasoning extraction. Some reasoning parsers
can legally consume the full output and return `(reasoning, None)`,
which makes the assert reachable and can surface as a server-side
failure.
This PR follows the same compatibility-patch pattern used by:
- `7314bbe2` fix(platform): reimplement MiniMax usage accounting patch
(#7835)
- `f83cb0e6` [Bugfix][Platform] Fix GLM47 tool-call finish backfill
(#7710)
The patch is intentionally narrow:
- normalize `content=None` to `""` only for forced named tool choice
- patch both chat-completions and responses parser entry points
- keep the rest of upstream behavior unchanged
Upstream tracking:
- issue: vllm-project/vllm#40147
- PR: vllm-project/vllm#40148
### Does this PR introduce _any_ user-facing change?
Yes.
Forced named tool choice becomes robust when the reasoning parser
returns no post-reasoning content, avoiding an internal assertion
failure and emitting an empty-argument function call instead.
### How was this patch tested?
Unit tests:
```bash
pytest -sv tests/ut/patch/platform/test_patch_tool_choice_none_content.py \
tests/ut/patch/platform/test_patch_glm_tool_call_parser.py \
tests/ut/patch/platform/test_patch_minimax_usage_accounting.py
```
Result: 22 passed.
---------
Signed-off-by: QwertyJack <7554089+QwertyJack@users.noreply.github.com>
Co-authored-by: QwertyJack <7554089+QwertyJack@users.noreply.github.com>
### What this PR does / why we need it?
This PR renames the environment variable VLLM_NIXL_ABORT_REQUEST_TIMEOUT
to VLLM_MOONCAKE_ABORT_REQUEST_TIMEOUT to align with the Mooncake
connector naming convention. It also updates the documentation and test
configurations to reflect this change and adjusts the suggested timeout
value in the documentation to 480 seconds for consistency.
### Does this PR introduce _any_ user-facing change?
Yes. The environment variable for configuring the abort request timeout
has been renamed. Users should update their environment settings from
VLLM_NIXL_ABORT_REQUEST_TIMEOUT to VLLM_MOONCAKE_ABORT_REQUEST_TIMEOUT.
### How was this patch tested?
The changes were verified by updating the corresponding test
configuration files and ensuring consistency across the documentation.
---------
Signed-off-by: herizhen <1270637059@qq.com>
Signed-off-by: herizhen <59841270+herizhen@users.noreply.github.com>
### What this PR does / why we need it?
This PR introduce stricter Ascend `additional_config.layer_sharding`
validation to the 0.18 release branch so it is only accepted on
PD-disaggregated P nodes with `kv_role="kv_producer"`.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
E2E test
---------
Signed-off-by: chenchuw886 <chenchuw@huawei.com>
Co-authored-by: chenchuw886 <chenchuw@huawei.com>
### What this PR does / why we need it?
Remove unused layers assignment in mooncake connector
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
by nightly
Signed-off-by: liziyu <liziyu16@huawei.com>
### What this PR does / why we need it?
fix tl.extract_slice and tl.insert_slice to extract_slice and
insert_slice from torch_utils
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
---------
Signed-off-by: wangx700 <wangxin700@huawei.com>
### What this PR does / why we need it?
Fix the issue where the Mooncake connector does not handle the MTP layer
KV cache when TP is unbalanced.
backport: #8540
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
by nightly
Signed-off-by: liziyu <liziyu16@huawei.com>
### What this PR does / why we need it?
Fix the issue where a request does not return due to a specific NPU on
node D having no transmission tasks in the scenario where node D is
enabled with DCP.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
by nightly
Signed-off-by: liziyu <liziyu16@huawei.com>
### What this PR does / why we need it?
This PR enables synchronization for the `PIECEWISE` runtime mode in ACL
graph replay. Previously, synchronization was only performed in `FULL`
mode. However, `PIECEWISE` mode also requires this barrier to ensure
that parameter updates are completed before the graph is replayed,
preventing accuracy loss.
The logic is also corrected to skip synchronization specifically for
EAGLE draft models, as intended.
Fixes #
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
CI passed.
---------
Signed-off-by: 1zzk <785396250@qq.com>
What this PR does / why we need it?
switch Ascend conv3d forward_oot to use forward_native and add ut
Does this PR introduce any user-facing change?
No
How was this patch tested?
by CI
---------
Signed-off-by: zouyizhou <zouyizhou@huawei.com>
### What this PR does / why we need it?
GDN Attention uses FIA's query_start_loc (padded), which may cause
conv1d update errors under high concurrency when dp > 1, and this PR is
to make GDN use its own query_start_loc (unpadded).
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
- vLLM version: v0.18.0
Signed-off-by: Wangbingjie <wangbj1207@126.com>
Reverts vllm-project/vllm-ascend#8133
- Reversion of Logic: This pull request reverts the changes introduced
in a previous commit that attempted to handle dimension mismatches
during SP padding.
Signed-off-by: Wangbingjie <wangbj1207@126.com>
<!-- Thanks for sending a pull request!
BEFORE SUBMITTING, PLEASE READ
https://docs.vllm.ai/en/latest/contributing/overview.html
-->
### What this PR does / why we need it?
- Enforce recompute scheduler only in PD-disaggregated mode.
- Enforce balance scheduling only in PD-mixed mode.
- Enforce fused MC2 only on PD-disaggregated D-side (kv_consumer).
<!--
- Please clarify what changes you are proposing. The purpose of this
section is to outline the changes and how this PR fixes the issue.
If possible, please consider writing useful notes for better and faster
reviews in your PR.
- Please clarify why the changes are needed. For instance, the use case
and bug description.
- Fixes #
-->
### Does this PR introduce _any_ user-facing change?
No
<!--
Note that it means *any* user-facing change including all aspects such
as API, interface or other behavior changes.
Documentation-only updates are not considered user-facing changes.
-->
### How was this patch tested?
By ci
<!--
CI passed with new added/existing test.
If it was tested in a way different from regular unit tests, please
clarify how you tested step by step, ideally copy and paste-able, so
that other reviewers can test and check, and descendants can verify in
the future.
If tests were not added, please describe why they were not added and/or
why it was difficult to add.
-->
---------
Signed-off-by: wangxiaoteng <wangxiaoteng@huawei.com>
### What this PR does / why we need it?
1. there is no synchronization between steps. However, in async
scheduling with aclgraph, it is possible that the CPU's record event for
the current iteration completes before the previous iteration's graph
execution has finished. If cpu is fast enough, device will hang on
event_wait in interation i+1 (assume that event_record is executed
immediately on update stream of device).
2. Under ENPU, eagle proposers also need to follow event.record first,
and then event.Wait.
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
---------
Signed-off-by: 1zzk <785396250@qq.com>
Cherry-picked from https://github.com/vllm-project/vllm-ascend/pull/7858
### What this PR does / why we need it?
This PR fixes a `RuntimeError` (dimension mismatch) that occurs when
Sequence Parallelism (SP) is enabled and the padding added for SP causes
`num_tokens_padded` to differ from `num_tokens_unpadded`. In such cases,
`_pad_query_start_loc_for_fia` adds a dummy request, increasing
`num_reqs_padded`. This mismatch between the actual number of requests
and the padded number of requests leads to errors in downstream token
count computations (e.g., `compute_num_computed_tokens`).
The fix modifies the restrictive condition `num_tokens_padded ==
num_tokens_unpadded` when reverting the dummy request padding if SP is
enabled, as SP padding is handled by stripping it after communication
and should not be treated as an additional request in the attention
metadata.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
vLLM version: v0.18.0
vLLM-Ascend version: releases/v0.18.0
Signed-off-by: Wangbj127 <wangbj1207@126.com>
### What this PR does / why we need it?
PR #8220 in v0.18.0
In a previous PR #7843 , the o_proj layer of GLM-5 was reverted to TP
(Tensor Parallel) splitting when flashcomm1 was enabled. However, this
was a temporary workaround and did not address the root cause of the
precision issues observed in the o_proj layer under flashcomm1.
I am working on a definitive fix for this issue. Currently, a clear bug
has been identified in
880e20fdde/vllm_ascend/quantization/methods/w8a8_static.py (L124):
during quantized matrix multiplication, quant_bias is not added if
tp_rank > 0. In the flashcomm1 scenario, all ranks actually require the
addition of quant_bias, meaning tp_rank=0 should be passed to ensure the
bias is applied correctly.
This PR aims to resolve this logic error and fix the underlying
precision issue.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
glm5 e2e test
---------
Signed-off-by: zjks98 <zhangjiakang4@huawei.com>
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: triomino <15924998+triomino@users.noreply.github.com>
Co-authored-by: zjks98 <zhangjiakang4@huawei.com>
### What this PR does / why we need it?
This PR is partially cherry-picked from #8172.
This PR aims to fix mismatched capture sizes after rounding operations
when using sp or speculative. The reason is that original
`self.cudagraph_capture_sizes` is no longer updated and remains as the
initial sizes. Now we use `self.cudagraph_dispatcher.get_capture_descs`
to the get up-to-date sizes.
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
by ci
Signed-off-by: Zetong Li <slippersss@126.com>
<!-- Thanks for sending a pull request!
BEFORE SUBMITTING, PLEASE READ
https://docs.vllm.ai/en/latest/contributing/overview.html
-->
### What this PR does / why we need it?
## Problem
In PD-disaggregated serving with `mooncake_connector` and
`VLLM_ASCEND_BALANCE_SCHEDULING=1`, requests may enter
`WAITING_FOR_REMOTE_KVS` and never be promoted back to runnable state
after remote KV transfer finishes.
The issue is in `BalanceScheduler`'s handling of
`WAITING_FOR_REMOTE_KVS` requests. The current code treats
`_update_waiting_for_remote_kv()` as if it returns a boolean readiness
flag:
```python
is_ready = self._update_waiting_for_remote_kv(request)
if is_ready:
...
else:
...
```
### Does this PR introduce _any_ user-facing change?
<!--
Note that it means *any* user-facing change including all aspects such
as API, interface or other behavior changes.
Documentation-only updates are not considered user-facing changes.
-->
### How was this patch tested?
<!--
CI passed with new added/existing test.
If it was tested in a way different from regular unit tests, please
clarify how you tested step by step, ideally copy and paste-able, so
that other reviewers can test and check, and descendants can verify in
the future.
If tests were not added, please describe why they were not added and/or
why it was difficult to add.
-->
Signed-off-by: Pz1116 <zpbzpb123123@gmail.com>
### What this PR does / why we need it?
This PR backports the DSA-CP PD role gating fix to `releases/v0.18.0`.
The existing helper logic on the release branch does not handle the PD
mixed-role case correctly when deciding whether layer sharding or TP
`o_proj` handling should be enabled. Layer sharding should only run on
the P-side instance, while TP `o_proj` handling should stay enabled for
normal non-PD deployments and for the PD mixed-role (`kv_both`)
instance. This patch makes those conditions explicit and adds unit
coverage for the allowed and disallowed combinations, including the
DSA-CP-disabled path.
Such wrong condition lead to **vllm serve failures** in case: **FC1 +
PD-colocated KV pooling + no layer_sharding**, specifically causing:
1. insufficient Available KV cache memory
2. o_proj shape error in sfa_v1 attention module
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
E2E test with dsv32 + FC1 + FULL_DECODE_ONLY +
kv_transfer_config(kv_both) + no layer_sharding
---------
Signed-off-by: chenchuw886 <chenchuw@huawei.com>
Co-authored-by: chenchuw886 <chenchuw@huawei.com>
### What this PR does / why we need it?
This PR implements the `AscendW8A8DynamicLinearMethod310` quantization
scheme specifically for 310P hardware. It includes the logic for weight
retrieval, per-channel parameter generation, and the application of
dynamic quantization using NPU-specific kernels. Additionally, it
updates `ShardedStateLoader310` to handle quantization configurations
more robustly when generating parameter type maps.
Feedback from the review identified two critical issues in the
implementation:
1. The tensor squeezing logic in the `apply` method incorrectly handles
2D inputs, which may lead to shape mismatches in subsequent layers.
2. The weight tensor in `process_weights_after_loading` is transposed
after being converted to the private NZ format; the transpose operation
should be performed on the ND tensor before conversion to ensure correct
physical layout.
cherry-pick from : #7546#7725
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
New unit tests were added in
`tests/ut/_310p/quantization/test_w8a8_dynamic_310.py` to verify the
quantization method, and
`tests/ut/_310p/test_sharded_state_loader_310p.py` was updated to test
the state loader changes.
---------
Signed-off-by: csoulnd <daidaicurry@foxmail.com>
<!-- Thanks for sending a pull request!
BEFORE SUBMITTING, PLEASE READ
https://docs.vllm.ai/en/latest/contributing/overview.html
-->
### What this PR does / why we need it?
<!--
- Please clarify what changes you are proposing. The purpose of this
section is to outline the changes and how this PR fixes the issue.
If possible, please consider writing useful notes for better and faster
reviews in your PR.
- Please clarify why the changes are needed. For instance, the use case
and bug description.
- Fixes #
-->
For the ENPU scenario, it is required that device events follow the
principle of "record first, wait later", otherwise the inference process
may become stuck. However, in the current model_forward function,
event.wait precedes event.record. Therefore, for the ENPU scenario,
graph parameter updates should be performed before model execution.
### Does this PR introduce _any_ user-facing change?
<!--
Note that it means *any* user-facing change including all aspects such
as API, interface or other behavior changes.
Documentation-only updates are not considered user-facing changes.
-->
N/A
### How was this patch tested?
<!--
CI passed with new added/existing test.
If it was tested in a way different from regular unit tests, please
clarify how you tested step by step, ideally copy and paste-able, so
that other reviewers can test and check, and descendants can verify in
the future.
If tests were not added, please describe why they were not added and/or
why it was difficult to add.
-->
---------
Signed-off-by: 1zzk <785396250@qq.com>
Signed-off-by: 1kzk <785396250@qq.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
### What this PR does / why we need it?
This PR backports the CPU binding locale normalization fix from #7274 to
`releases/v0.18.0`, including the follow-up review fixes already applied
on `main`.
The change forces `LC_ALL`, `LANG`, and `LC_MESSAGES` to `C` before
spawning subprocesses in `vllm_ascend.cpu_binding.execute_command()`, so
parser-dependent command output stays stable on localized systems. It
also handles `subprocess.TimeoutExpired` by killing the child process
before collecting output, and updates the existing unit tests to keep
command-argument coverage while adding timeout-path coverage.
Fixes#6992
### Does this PR introduce _any_ user-facing change?
Yes.
Users running CPU binding on non-English OS environments should now get
consistent English subprocess output for parser-dependent commands,
avoiding failures caused by inherited locale settings.
### How was this patch tested?
- Updated the existing unit tests in
`tests/ut/device_allocator/test_cpu_binding.py` to assert the locale
environment, retain command argument coverage, and cover the timeout
cleanup path.
- Attempted to run targeted pytest cases locally, but the pytest
invocation did not complete normally in this environment, so I could not
record a clean passing run here.
Attribution:
- Co-authored-by: stdjhs <1601599324@qq.com>
- Signed-off-by: chenchuw886 <chenchuw@huawei.com>
Signed-off-by: chenchuw886 <chenchuw@huawei.com>
Co-authored-by: chenchuw886 <chenchuw@huawei.com>
Co-authored-by: stdjhs <1601599324@qq.com>
<!-- Thanks for sending a pull request!
BEFORE SUBMITTING, PLEASE READ
https://docs.vllm.ai/en/latest/contributing/overview.html
-->
### What this PR does / why we need it?
ref:https://github.com/vllm-project/vllm-ascend/issues/8184
following https://github.com/vllm-project/vllm/pull/31057, add
`requires_piecewise_for_cudagraph` for `AscendStoreConnector`
### Does this PR introduce _any_ user-facing change?
<!--
Note that it means *any* user-facing change including all aspects such
as API, interface or other behavior changes.
Documentation-only updates are not considered user-facing changes.
-->
### How was this patch tested?
<!--
CI passed with new added/existing test.
If it was tested in a way different from regular unit tests, please
clarify how you tested step by step, ideally copy and paste-able, so
that other reviewers can test and check, and descendants can verify in
the future.
If tests were not added, please describe why they were not added and/or
why it was difficult to add.
-->
Signed-off-by: Pz1116 <zpbzpb123123@gmail.com>
### What this PR does / why we need it?
Fix a bug in the GLM tool call parser where the `function.name` field
was incorrectly included in the final (non-first) chunks of streaming
tool calls.
Per OpenAI streaming semantics, `id`, `type`, and `function.name` must
only appear in the **first** chunk for a given tool call index. When
`_create_remaining_args_delta` was called for continuing/finishing
chunks, it was incorrectly reading the function name from
`delta_message.tool_calls` and re-emitting it, causing clients to see a
duplicate/extra function name in the final chunk.
**Root cause**: The original code always looked up the tool call in
`delta_message.tool_calls` to get the name, id, and type — even when
this was not the first chunk being streamed. This caused the function
name to appear again in the final argument-completion chunk.
**Fix**:
- Track whether arguments have already been streamed
(`already_streamed_args`) for each tool call index.
- Only populate `fallback_tool_call_id`, `fallback_tool_call_type`, and
`fallback_tool_call_name` when `already_streamed_args` is empty (i.e.,
this is genuinely the first chunk).
- Refactored `_create_remaining_args_delta` to omit header fields
entirely when all fallback values are `None`, which is the correct
behavior for continuing/finishing chunks.
### Does this PR introduce _any_ user-facing change?
Yes. Clients consuming the streaming tool call response will no longer
receive a duplicate `function.name` in the final chunk. This fixes
incorrect behavior visible in the OpenAI-compatible streaming API output
for GLM models using tool calls.
### How was this patch tested?
- Code review and logic analysis of the streaming tool call path in
`patch_glm_tool_call_parser.py`.
- Existing unit tests in
`tests/ut/platform/test_patch_glm_tool_call_parser.py`.
---------
Signed-off-by: chen-weipeng12 <chen-weipeng12@noreply.gitcode.com>
Signed-off-by: chenweiqiang11 <chenweiqiang11@noreply.github.com>
Co-authored-by: chen-weipeng12 <chen-weipeng12@noreply.gitcode.com>
### What this PR does / why we need it?
This PR is cherry-pick from #8263.
This PR aims to fix short prompt problem. The root cause can be found in
#8029. Since the previous pr may miss mixed long and short prompt batch,
after discussion, we decide to add PrefillNoCache state in mla
_forward_decode now instead.
Signed-off-by: Zetong Li <slippersss@126.com>
### What this PR does / why we need it?
This PR improves the validation of `max_cudagraph_capture_size` by
comparing it against the potential maximum tokens required for decoding,
derived from the scheduler configuration. It introduces a warning to
alert users when the capture size might be insufficient for the
workload, which could lead to suboptimal performance.
ref: #8227
### Does this PR introduce _any_ user-facing change?
Yes, a warning log is added when the `max_cudagraph_capture_size` is
smaller than the potential decode workload.
---------
Signed-off-by: QiuChunshuo <qiuchunshuo@huawei.com>
### What this PR does / why we need it?
This PR fixes a service startup failure for DeepSeek-V3.1 models by
removing a strict type assertion for `MLAAttentionSpec` in
`NPUModelRunner.get_kv_cache_spec`. The assertion was failing due to
class identity mismatches caused by the runtime patching of
`MLAAttentionSpec` with `AscendMLAAttentionSpec`.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Verified that the service starts correctly for DSV3.1 models.
Signed-off-by: mayumeng <m30059191@china.huawei.com>
Co-authored-by: mayumeng <m30059191@china.huawei.com>
### What this PR does / why we need it?
Fix Qwen3.5 MoE MTP layer shared expert shape error when flash comm v1
is enabled.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
- vLLM version: v0.18.0
- vLLM main:
35141a7eed
Signed-off-by: Wangbingjie <wangbj1207@126.com>
### What this PR does / why we need it?
bugfix short squence has no respone. This pull request refactors the
event handling for KV cache reshaping in mla_v1.py by centralizing the
reshape_cache_event creation and recording within the _mla_preprocess
function, ensuring it covers both decode and prefill operations.
Signed-off-by: wangxiaoteng <wangxiaoteng@huawei.com>
### What this PR does / why we need it?
Introduce a check to not using asynchronous communication under
`enable_dsa_cp_with_layer_shard` branch on capturing mode. This change
prevents potential stream and event issues when operating in
graph/capturing mode, ensuring safer communication practices.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
E2E test with dsv32 + FC1 + FULL_DECODE_ONLY +
kv_transfer_config(kv_both)
---------
Signed-off-by: chenchuw886 <chenchuw@huawei.com>
Co-authored-by: chenchuw886 <chenchuw@huawei.com>
…and PCP across machines
<!-- Thanks for sending a pull request!
BEFORE SUBMITTING, PLEASE READ
https://docs.vllm.ai/en/latest/contributing/overview.html
-->
### What this PR does / why we need it?
<!--
- Please clarify what changes you are proposing. The purpose of this
section is to outline the changes and how this PR fixes the issue.
If possible, please consider writing useful notes for better and faster
reviews in your PR.
- Please clarify why the changes are needed. For instance, the use case
and bug description.
- Fixes #
-->
### Does this PR introduce _any_ user-facing change?
<!--
Note that it means *any* user-facing change including all aspects such
as API, interface or other behavior changes.
Documentation-only updates are not considered user-facing changes.
-->
### How was this patch tested?
<!--
CI passed with new added/existing test.
If it was tested in a way different from regular unit tests, please
clarify how you tested step by step, ideally copy and paste-able, so
that other reviewers can test and check, and descendants can verify in
the future.
If tests were not added, please describe why they were not added and/or
why it was difficult to add.
-->
Signed-off-by: DreamLeader <2270923832@qq.com>
### What this PR does / why we need it?
This PR is cherry-pick from #8029.
This PR aims to fix attention state of short prompt for correct
forwarding. Since a batch of short prompts (prefill tokens less than or
equal to num_spec_tokens + 1) will be treated as decode requests (by
split_decodes_and_prefills), its original PrefillNoCache attention state
contradicts. Thus these short prompts will be passed into a mismatched
branch and incur errors.
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
by ci
Signed-off-by: Zetong Li <slippersss@126.com>
This pull request reverts previous changes to switch to FIA and instead
implements npu_ring_mla for MLA prefill operations(#5704 ). The change
streamlines the attention mechanism by removing unnecessary metadata
tracking and updating the underlying NPU operations to use the
ring-based MLA kernel. This adjustment ensures better compatibility and
performance for MLA prefill tasks within the vLLM Ascend backend.
Highlights
- Migration to npu_ring_mla: Replaced the usage of
npu_fused_infer_attention_score (FIA) with npu_ring_mla for MLA prefill
operations across the codebase to improve performance and alignment with
the intended architecture.
- Cleanup of redundant metadata: Removed
chunk_actual_seq_lengths_kv_list and actual_seq_lengths_q from various
metadata structures as they are no longer required for the updated
attention implementation.
- Test suite updates: Updated unit tests in test_mla_cp.py and
test_mla_v1.py to mock npu_ring_mla instead of the deprecated FIA
functions and adjusted test assertions to reflect the new implementation
details.
Signed-off-by: weijinqian_v1 <weijinqian@huawei.com>
Co-authored-by: weijinqian_v1 <weijinqian@huawei.com>
### What this PR does / why we need it?
Cherry-picked from #8039
Restore the setting of MC2 `global_bs` and `mc2_mask` handling when
`all_reduce` across DP group cannot be skipped. Ascend MC2 ops require
`global_bs=0` + `mc2_mask` while enabling inter-node roce hierarchical
communication. PR #4983 always passed non-zero `global_bs` without
`mc2_mask`, which is incompatible with hierarchy comm raised in PR #7583
**Changes:**
- Add `should_skip_allreduce_across_dp_group()` to `utils.py` with
hierarchy constraint
- Set `global_bs=0` when allreduce is not skipped; pass `mc2_mask`
accordingly
- Add `mc2_mask` field to `MoEMC2CombineMetadata` for dispatch→combine
propagation
### Does this PR introduce _any_ user-facing change?
No. But this PR fixes cross-super-node communication function on A3 with
`enable_mc2_hierarchy_comm=True` in `additional_config` and `export
HCCL_INTRA_ROCE_ENABLE=1`.
### How was this patch tested?
E2E serving succeeded and CI pssed.
- vLLM version: v0.18.0
- vLLM main:
14acf429ac
---------
Signed-off-by: linfeng-yuan <1102311262@qq.com>
### What this PR does / why we need it?
Enabling temperature in post-processing on 310P devices can cause the
service to stall and eventually hang. We first traced the issue to a
timeout where the temperature-related `div` operator was waiting for
results from a sub-stream. After investigating the preceding operators,
we finally identified the root cause as the `q.exponential_()` operator,
which is not well supported on 310P and triggers an internal issue in
the `add` kernel.
### Does this PR introduce _any_ user-facing change?
NA
### How was this patch tested?
This patch was thoroughly tested locally（accuracy-dataset test and
stress test）. It is not easy to design a proper unit test for this case,
and I appreciate your understanding.
Signed-off-by: Tflowers-0129 <2906339855@qq.com>
<!-- Thanks for sending a pull request!
BEFORE SUBMITTING, PLEASE READ
https://docs.vllm.ai/en/latest/contributing/overview.html
-->
### What this PR does / why we need it?
When we launch a PD-disaggregated process and send requests, an
additional processes appear on NPU 0, becasue when a thread has a
primary cuda context, the child thread it creates automatically doesn't
inherit the cuda context. See
https://forums.developer.nvidia.com/t/when-a-thread-has-a-primary-cuda-context-does-the-child-thread-it-creates-automatically-inherit-the-cuda-context/362810.
vLLM has fixed this issue in [pr-37449
](https://github.com/vllm-project/vllm/pull/37449), but version 0.18.0
does not include the fix. Therefore, we need to patch it.
<!--
- Please clarify what changes you are proposing. The purpose of this
section is to outline the changes and how this PR fixes the issue.
If possible, please consider writing useful notes for better and faster
reviews in your PR.
- Please clarify why the changes are needed. For instance, the use case
and bug description.
- Fixes #
-->
### Does this PR introduce _any_ user-facing change?
no
<!--
Note that it means *any* user-facing change including all aspects such
as API, interface or other behavior changes.
Documentation-only updates are not considered user-facing changes.
-->
### How was this patch tested?
<!--
CI passed with new added/existing test.
If it was tested in a way different from regular unit tests, please
clarify how you tested step by step, ideally copy and paste-able, so
that other reviewers can test and check, and descendants can verify in
the future.
If tests were not added, please describe why they were not added and/or
why it was difficult to add.
-->
---------
Signed-off-by: zouyida <zouyida@huawei.com>
Co-authored-by: zouyida <zouyida@huawei.com>
backport of #7474
This PR adds C8 (INT8) KV cache quantization support for standard GQA
attention models (e.g., Qwen3-32B W8A8C8). C8 uses static per-channel
quantization scales to store KV cache in INT8, reducing KV cache memory
by ~50% compared to BF16, enabling higher batch concurrency and longer
context lengths on the same hardware.
**Key changes:**
1. **`attention_v1.py`** — New `AscendC8AttentionBackendImpl` subclass
of `AscendAttentionBackendImpl`:
- `_prepare_c8_scales`: Shards per-channel scales/offsets to the current
TP rank and pre-computes BF16 BNSD-shaped antiquant tensors (one-time
per layer).
- `_quantize_kv_to_int8`: Quantizes BF16 K/V to INT8 before
`reshape_and_cache`, using pre-cached inverse scales.
- `_forward_c8_decode`: FIA V1 BNSD paged attention with native INT8 KV
and `perchannel` antiquant mode.
- `_forward_c8_chunked_prefill`: Splits decode (FIA V1 BNSD paged INT8)
and prefill (FIA V1 TND float) into two kernel calls.
- `_forward_c8_fused_infer_attention`: Handles `PrefillNoCache` and
`PrefillCacheHit` states.
2. **`quantization/methods/kv_c8.py`** — New
`AscendC8KVCacheAttentionMethod` scheme:
- Creates `k/v_cache_scale/offset` parameters via
`_c8_kv_scale_weight_loader`, which handles per-channel scale shapes and
lazy resizing.
- Sets `layer.kv_cache_torch_dtype = torch.int8` so
`get_kv_cache_spec()` returns INT8 dtype automatically.
- Upgrades `layer.impl` to `AscendC8AttentionBackendImpl` via class
surgery.
3. **`quantization/modelslim_config.py`** — C8 branch in
`get_quant_method()` activates when `kv_cache_type == "C8"` in
`quant_model_description.json`.
4. **`patch/worker/patch_qwen3_c8.py`** — Intercepts per-channel C8
scale/offset weights before `AutoWeightsLoader` discards them, routing
them to the parameters created by `AscendC8KVCacheAttentionMethod`.
5. **`tests/ut/quantization/test_kv_c8.py`** — Unit tests covering
`_c8_kv_scale_weight_loader`, `AscendC8KVCacheAttentionMethod`, and
`AscendC8AttentionBackendImpl` scale helpers.
Yes. Users can now serve Qwen3-32B W8A8C8 quantized models with INT8 KV
cache on Ascend NPU. The model checkpoint must contain a
`quant_model_description.json` with `"kv_cache_type": "C8"` and
per-channel scale/offset tensors in safetensors.
No changes to the serving CLI — the feature activates automatically when
the quantization config is detected.
Benchmarked with `vllm serve` (TP=8, `max_num_seqs=256`,
`max_model_len=131072`, `enable_chunked_prefill=true`) + `random_bench`
(input_len=10240, output_len=2048, 960 prompts, max_concurrency=192):
```
============ Serving Benchmark Result ============
Successful requests: 960
Failed requests: 0
Maximum request concurrency: 192
Benchmark duration (s): 1359.81
Total input tokens: 9830400
Total generated tokens: 1966080
Request throughput (req/s): 0.71
Output token throughput (tok/s): 1445.85
Peak output token throughput (tok/s): 2304.00
Total token throughput (tok/s): 8675.12
---------------Time to First Token----------------
Mean TTFT (ms): 24598.51
Median TTFT (ms): 23167.02
P50 TTFT (ms): 23167.02
P90 TTFT (ms): 47717.08
P99 TTFT (ms): 84402.61
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms): 120.76
Median TPOT (ms): 121.50
P50 TPOT (ms): 121.50
P90 TPOT (ms): 127.05
P99 TPOT (ms): 130.13
---------------Inter-token Latency----------------
Mean ITL (ms): 120.70
Median ITL (ms): 90.34
P50 ITL (ms): 90.34
P90 ITL (ms): 93.79
P99 ITL (ms): 101.80
==================================================
```
All attention states verified: `PrefillNoCache`, `PrefillCacheHit`,
`ChunkedPrefill`, `DecodeOnly`.
- vLLM version: v0.17.0
- vLLM main:
8b6325758c
Signed-off-by: lico67373 <918688502@qq.com>
Co-authored-by: LICO67373 <110013619+LICO1314@users.noreply.github.com>
### What this PR does / why we need it?
Enable Flash Comm V1 (sequence parallelism) for Qwen3-VL models (both
dense and MoE variants).
Root cause: Qwen3-VL's deepstack embeddings remain full-size [N, H]
while hidden states become [N/tp_size, H] after reduce-scatter, causing
shape mismatch on add.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- [x] Run Qwen3-VL dense model with FC1 enabled (TP > 1), verify correct
output
- [x] Run Qwen3-VL MoE model with FC1 enabled (TP > 1), verify correct
output
---------
Signed-off-by: betta18 <jiangmengyu1@huawei.com>
Signed-off-by: jiangmengyu18 <56633611+jiangmengyu18@users.noreply.github.com>
Co-authored-by: betta18 <jiangmengyu1@huawei.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
### What this PR does / why we need it?
Qwen3vl full attention supports enabling the split_qkv_rmsnorm_mrope
fusion operator.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- [x] Run Qwen3-VL dense model with the fusion operator, verify correct
output
- [x] Run Qwen3-VL MoE model with the fusion operator, verify correct
output
---------
Signed-off-by: jiangmengyu18 <451528648@qq.com>
Signed-off-by: jiangmengyu18 <56633611+jiangmengyu18@users.noreply.github.com>
Signed-off-by: betta18 <jiangmengyu1@huawei.com>
Co-authored-by: betta18 <jiangmengyu1@huawei.com>
### What this PR does / why we need it?
This PR fixes a weight loading error in the Qwen3-VL model.
The bug was introduced by vLLM. In vLLM's `qwen3-vl.py`, the prefix of
the `lm_head` layer is hardcoded as `"lm_head"`. However,
`hf_to_vllm_mapper` remaps the weight name of `lm_head` from `lm_head`
to `language_model.lm_head`.
This causes a mismatch between the keys in the weight file and the
prefix of the lm_head layer, resulting in an error.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- [x] Run Qwen3-VL dense model with the fusion operator, verify correct
output
Signed-off-by: betta18 <jiangmengyu1@huawei.com>
Co-authored-by: betta18 <jiangmengyu1@huawei.com>
<!-- Thanks for sending a pull request!
BEFORE SUBMITTING, PLEASE READ
https://docs.vllm.ai/en/latest/contributing/overview.html
-->
### What this PR does / why we need it?
vLLM v0.18 defers KV connector finalization during target-modelforward
when speculative decoding is enable, leading to KV Pool not doing Put
Operation. This change is forgotten when we bumpped up the version for
vllm-ascend. Fix by adding finalize_kv_connector for spec decode.
<!--
- Please clarify what changes you are proposing. The purpose of this
section is to outline the changes and how this PR fixes the issue.
If possible, please consider writing useful notes for better and faster
reviews in your PR.
- Please clarify why the changes are needed. For instance, the use case
and bug description.
- Fixes #
-->
### Does this PR introduce _any_ user-facing change?
<!--
Note that it means *any* user-facing change including all aspects such
as API, interface or other behavior changes.
Documentation-only updates are not considered user-facing changes.
-->
### How was this patch tested?
<!--
CI passed with new added/existing test.
If it was tested in a way different from regular unit tests, please
clarify how you tested step by step, ideally copy and paste-able, so
that other reviewers can test and check, and descendants can verify in
the future.
If tests were not added, please describe why they were not added and/or
why it was difficult to add.
-->
Signed-off-by: Pz1116 <zpbzpb123123@gmail.com>
Co-authored-by: DreamerLeader <2270923832@qq.com>
Co-authored-by: fems14 <1804143737@qq.com>
