### 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 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?
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>
<!-- 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?
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>
<!-- 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>
## Summary
- replace the MiniMax usage accounting monkey patch with a runtime
wrapper implementation instead of source-text rewriting
- preserve MiniMax reasoning-token semantics when `</think>` is missing
by counting the emitted output as reasoning tokens
- add unit coverage for usage tracking helpers and MiniMax
reasoning-token counting
## Why
The previous implementation rewrote `OpenAIServingChat` by matching
exact source blocks. That was brittle against `vllm` source drift and
could crash during early plugin initialization with:
`RuntimeError: Failed to locate expected block while patching
OpenAIServingChat usage accounting.`
This change keeps the usage-accounting backport, but applies it by
wrapping the original stream/full generators and tracking output token
ids at runtime.
For MiniMax reasoning counting, a missing `</think>` should not be
treated as zero reasoning tokens. It can mean the whole output is still
in thinking mode, or that generation stopped before the closing token
was produced. In that case, the emitted output should still be counted
as reasoning.
## Validation
- `pytest -q
tests/ut/patch/platform/test_patch_minimax_usage_accounting.py`
- `vllm serve --help`
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?
Optimize DeepSeekOCR2 RelPosAttention and CustomQwen2Decoder and add doc
for DeepSeekOCR2.md
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vllm 0.18.0
- vllm-ascend main
1. _create_custom_4d_mask during 141ms49us620ns -->
_create_npu_optimized_mask during 1ms227us780ns
2. convd2d : 27ms --> matmul <1ms
3. relposattention：sdpa->prompt_flash_attention
---------
Signed-off-by: Wangbei25 <wangbei41@huawie.com>
Signed-off-by: Wangbei25 <wangbei41@huawei.com>
Co-authored-by: Wangbei25 <wangbei41@huawie.com>
### What this PR does / why we need it?
This rebases the GLM47 tool-call parser fix onto `releases/v0.18.0`
after the MiniMax usage-accounting patch merged upstream on March 27,
2026.
It fixes OpenAI chat tool-call streaming for GLM47 by:
- draining terminal parser chunks that contain both the final argument
text and the closing `</tool_call>` suffix
- computing finish backfill from the tool argument bytes actually
emitted to the client, instead of trusting parser-internal buffered
state
- adding focused regression tests for finish backfill and terminal chunk
handling
### Does this PR introduce _any_ user-facing change?
Yes. GLM47 OpenAI-compatible streaming tool-call responses now emit
correct final chunks and argument payloads on `releases/v0.18.0`.
### How was this patch tested?
- `pytest -q tests/ut/patch/platform/test_patch_glm_tool_call_parser.py
tests/ut/patch/platform/test_patch_minimax_usage_accounting.py`
- `python -m pre_commit run --files
vllm_ascend/patch/platform/patch_glm_tool_call_parser.py
tests/ut/patch/platform/test_patch_glm_tool_call_parser.py
vllm_ascend/patch/platform/__init__.py vllm_ascend/patch/__init__.py`
---------
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 backports the MiniMax M2 reasoning-token usage accounting fix onto
`releases/v0.18.0` for vllm-ascend.
The release branch does not include the other local GLM patch commit, so
this PR keeps the MiniMax change self-contained by:
- registering `patch_minimax_usage_accounting` on the release branch
- backporting `completion_tokens_details.reasoning_tokens` into chat
usage generation
- fixing MiniMax reasoning token counting for `</think>`-delimited
outputs without depending on the GLM suffix patch
### Does this PR introduce _any_ user-facing change?
Yes. OpenAI-compatible chat usage accounting for MiniMax M2 responses
now reports corrected reasoning token counts on the release branch.
### How was this patch tested?
- `python -m compileall
vllm_ascend/patch/platform/patch_minimax_usage_accounting.py`
- `python - <<'PY'` import check for
`vllm_ascend.patch.platform.patch_minimax_usage_accounting` on top of
`releases/v0.18.0`
No targeted automated regression test exists for this release-branch
backport yet, so I validated syntax and module import compatibility on
the release branch.
---------
Signed-off-by: QwertyJack <7554089+QwertyJack@users.noreply.github.com>
Co-authored-by: QwertyJack <7554089+QwertyJack@users.noreply.github.com>
cherry pick from https://github.com/vllm-project/vllm-ascend/pull/7486
<!-- 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 #
-->
Multimodal models like Qwen3.5 MoE does embedding in model_runner, so
when flash comm is enabled, the first AllGather operation should be
skipped.
### 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.
-->
No.
### 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.
-->
- vLLM version: v0.18.0
- vLLM main:
8b6325758c
---------
Signed-off-by: Wangbingjie <wangbj1207@126.com>
Signed-off-by: wangbj127 <256472688+wangbj127@users.noreply.github.com>
### What this PR does / why we need it?
During the attention quantization process of DeepSeek V3.2, it is
necessary to retrieve the Hadamard matrix from the weights to facilitate
the computation.
### Does this PR introduce _any_ user-facing change?
No. But there will be two new tensor in quant weight.
### How was this patch tested?
- vLLM version: v0.18.0
- vLLM main:
8b6325758c
---------
Signed-off-by: mayumeng <m30059191@china.huawei.com>
Co-authored-by: mayumeng <m30059191@china.huawei.com>
### What this PR does / why we need it?
This PR aims to adapt to newest commit of vllm main branch for model
runner v2. please refer to
https://github.com/vllm-project/vllm-ascend/issues/5208
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
- vLLM version: v0.18.0
- vLLM main:
ed359c497a
---------
Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
### What this PR does / why we need it?
This PR introduces a "balance scheduling" feature, enabled by the
`VLLM_ASCEND_BALANCE_SCHEDULING` environment variable. This feature
adjusts the scheduling logic to better balance the load across
data-parallel workers, preventing a single worker from blocking
scheduling for others. This can improve overall throughput.
Additionally, this PR includes a number of other updates and fixes to
the scheduler, syncing it with a more recent version of the upstream
vLLM scheduler. These changes include:
- Handling for paused scheduler state.
- Support for Mamba block-aligned splits.
- Handling for streaming requests.
- Refinements in preemption logic and resource management (KV cache,
encoder cache).
- General code refactoring for clarity and correctness.
Fixes #
### Does this PR introduce _any_ user-facing change?
Yes, this PR introduces a new feature controlled by the
`VLLM_ASCEND_BALANCE_SCHEDULING` environment variable. When enabled, the
scheduling behavior changes, which could affect performance and request
throughput.
### How was this patch tested?
CI passed. Further testing should be done to validate the performance
and correctness of the new scheduling logic under various workloads,
with and without the feature flag enabled.
Signed-off-by: GDzhu01 <809721801@qq.com>
### What this PR does / why we need it?
2nd PR for https://github.com/vllm-project/vllm-ascend/issues/5712,
extend SP to VL MoE models.
### Does this PR introduce _any_ user-facing change?
remove `sp_threshold` in additional config and reuse `sp_min_token_num`
from vLLM.
### How was this patch tested?
- Model: Qwen3-VL-30B-A3B,
- TP4 DP2
- 100 reqs
- max concurrency 1
| Seq length | Mean TTFT (ms) main | Mean TTFT (ms) this PR |
|------------|---------------------|------------------------|
| 4k | 429.40 | 323.3 |
| 16k | 1297.01 | 911.74 |
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: realliujiaxu <realliujiaxu@163.com>
### What this PR does / why we need it?
During the prefill phase of Qwen3-Next and Qwen3.5, the
`torch.ops._C_ascend.causal_conv1d_fn` operator exhibits significant
performance bottlenecks. To address this, we have re-implemented the
optimization using `torch.ops._C_ascend.npu_causal_conv1d_custom`.
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
1 accuracy test
```
[2026-03-20 16:44:22,961] [ais_bench] [INFO] Start launch task state board ...
+-----------------------------+-----------+------------+-------------+----------+-------------------------------------------+---------------------+
| Task Name | Process | Progress | Time Cost | Status | Log Path | Extend Parameters |
+=============================+===========+============+=============+==========+===========================================+=====================+
| vllm-api-general-chat/gsm8k | 2918978 | NA | 0:00:01 | finish | logs/eval/vllm-api-general-chat/gsm8k.out | None |
+-----------------------------+-----------+------------+-------------+----------+-------------------------------------------+---------------------+
[2026-03-20 16:44:34,284] [ais_bench] [INFO] Evaluation tasks completed.
[2026-03-20 16:44:34,287] [ais_bench] [INFO] Summarizing evaluation results...
dataset version metric mode vllm-api-general-chat
--------- --------- -------- ------ -----------------------
gsm8k 271d0b accuracy gen 96.21
```
2 ut modify test
`pytest -sv
/home/c30006096/vllm-ascend/tests/e2e/nightly/single_node/ops/singlecard_ops/triton/test_causal_conv1d.py::test_ascend_causal_conv1d`
- vLLM version: v0.17.0
- vLLM main:
8b6325758c
Signed-off-by: wenba0 <3054239545@qq.com>
Signed-off-by: jiaojiao <56385650+wenba0@users.noreply.github.com>
### What this PR does / why we need it?
Qwen3.5 full attention supports enabling the split_qkv_rmsnorm_mrope
fusion operator.
### How was this patch tested?
vLLM version: v0.16.0
vLLM-Ascend main: https://github.com/vllm-project/vllm-ascend/pull/6730
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: ZhuQi-seu <zhuqi12@huawei.com>
### What this PR does / why we need it?
Upgrade vllm commit to 2026.03.19.
1.Fix socket removed from StatelessProcessGroup. Upstream vLLM PR
[#36330](https://github.com/vllm-project/vllm/pull/36330) ("elastic_ep:
Fix stateless group port races") refactored StatelessProcessGroup and
removed the socket: socket.socket | None field. The socket ownership was
moved to a new create_tcp_store() helper instead of being stored as a
field on the dataclass.
2.fix `virtual_engine` parameter removed from `set_forward_context().
Upstream [V0 Deprecation] Deprecate virtual engine
[#37195](https://github.com/vllm-project/vllm/pull/37195)
### Does this PR introduce _any_ user-facing change?
NA
### How was this patch tested?
NA
- vLLM version: v0.17.0
- vLLM main:
8b6325758c
---------
Signed-off-by: leo-pony <nengjunma@outlook.com>
### What this PR does / why we need it?
This PR optimizes the Qwen3.5 and Qwen3Next GDN prefill path on Ascend
by reducing host/device synchronization overhead.
The current implementation of the `chunk_gated_delta_rule` path for
variable-length sequences prepares chunk metadata during the forward
pass. This approach triggers frequent CPU intervention and host/device
round-trips. When running prefill-heavy workloads with asynchronous
scheduling enabled, these synchronizations result in execution "bubbles"
and prefill stalling (stuttering). **Note that this does not cause
asynchronous scheduling to fail; rather, it prevents the system from
reaching its theoretical throughput due to these unnecessary stalls.**
To resolve this, the patch moves metadata preparation out of the hot
path:
- **Prebuilt Metadata:** All non-speculative varlen chunk metadata for
GDN is now prebuilt on the CPU.
- **Asynchronous Transfer:** Staging buffers are kept in pinned memory
and transferred to the NPU asynchronously.
- **Integration:** The prebuilt bundle is attached to GDN attention
metadata via `patch_gdn_attn.py` and passed into Triton wrappers.
- **Backward Compatibility:** Triton wrappers fall back to the legacy
preparation path if no prebuilt metadata is provided.
- vLLM version: v0.17.0
- vLLM main:
8b6325758c
---------
Signed-off-by: maoxx241 <maomaoyu870@gmail.com>
### What this PR does / why we need it?
Delete the logic that the input of get_rope_shape from device to host.
- vLLM version: v0.17.0
- vLLM main:
8b6325758c
Signed-off-by: LoganJane <loganJane73@hotmail.com>
### What this PR does / why we need it?
1. upgrade to 0.18.0
2. ensure kernel_block_sizes is int for Eagle drafter
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.17.0
- vLLM main:
8b6325758c
---------
Signed-off-by: Meihan-chen <jcccx.cmh@gmail.com>
Signed-off-by: hfadzxy <starmoon_zhang@163.com>
Co-authored-by: hfadzxy <starmoon_zhang@163.com>
### What this PR does / why we need it?
This PR introduces a new fused Triton kernel,
`split_qkv_tp_rmsnorm_rope` for Minimax-m2.5.
The implementation includes two Triton kernels:
1. `_split_qkv_and_compute_local_qk_var_kernel`: Splits the QKV input
and computes the local variance for RMSNorm.
2. `_apply_global_rmsnorm_kernel`: Applies global RMSNorm (considering
TP all-reduce for variance) and Neox-style RoPE.
### Does this PR introduce _any_ user-facing change?
Does not.
### How was this patch tested?
```python
pytest tests/e2e/nightly/single_node/ops/singlecard_ops/triton/test_split_qkv_tp_rmsnorm_rope.py
```
### Test Data
A3 TP16
基线
| data | TTFT(ms) | TPOT(ms) | TPS |
|------------|---------:|---------:|-------:|
| 4k/1k@bs1 | 267.55 | 25.5 | 38.85 |
| 4k/1k@bs4 | 542.4 | 26.51 | 148.06 |
测试线
| data | TTFT(ms) | TPOT(ms) | TPS |
|------------|---------:|---------:|-------:|
| 4k/1k@bs1 | 234.64 | 20.96 | 47.24 |
| 4k/1k@bs4 | 508.36 | 22.16 | 176.69 |
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
Signed-off-by: xutianyi <xutianyi5@huawei.com>
Co-authored-by: xutianyi <xutianyi5@huawei.com>
### What this PR does / why we need it?
1. Mamba Cache Support on 310P: Implemented logic to correctly
initialize and allocate KV cache for Mamba models on the 310P platform,
including handling of state tensors and page size alignment.
2. Increased Attention Head Size Support: Modified the attention backend
to support attn_head_size larger than 128 by dynamically selecting
appropriate kernel block sizes based on hardware limitations (e.g.,
block_size * head_size <= 16384).
3. Refactored KV Cache Allocation: Consolidated and improved the KV
cache allocation mechanism, moving from separate size calculation and
allocation steps to a unified _allocate_kv_cache_tensors method that
handles both Attention and Mamba specific cache structures.
4. Dynamic Mamba Config Patching: Introduced conditional loading of
Mamba configuration patches, specifically using patch_mamba_config_310
for the 310P platform to ensure platform-specific optimizations and
validations.
5. Reserve reasonable memory to allocate KV cache to avoid OOM issue
with default gpu_memory_utilization.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Qwen3.5 E2E test
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: pu-zhe <zpuaa@outlook.com>
### What this PR does / why we need it?
1.fix "TypeError: get_attn_backend() remove variable": [Refactor
`check_and_update_config`](https://github.com/vllm-project/vllm/pull/35122)
2.fix [Rename `compile_ranges_split_points` to
`compile_ranges_endpoints`](https://github.com/vllm-project/vllm/pull/36027)
3.fix "RuntimeError: device_allocator not a DeviceAllocator":[Replace
memory related torch.cuda
APIs"](https://github.com/vllm-project/vllm/pull/37031)
4.fix [Support multiple KV groups in OffloadingSpec
](https://github.com/vllm-project/vllm/pull/36610) removed
self.offloaded_block_size and changed self.gpu_block_size from a scalar
to a tuple of per-group block sizes, adding block_size_factor.
5.fix [Consolidate
SupportsEagle](https://github.com/vllm-project/vllm/pull/36063) renamed
get_eagle3_aux_hidden_state_layers() to
get_eagle3_default_aux_hidden_state_layers() and added a
supports_eagle3() guard before calling it.
### Does this PR introduce _any_ user-facing change?
NA
### How was this patch tested?
E2E
- vLLM version: v0.17.0
- vLLM main:
8a680463fa
---------
Signed-off-by: leo-pony <nengjunma@outlook.com>
Co-authored-by: Claude Code <noreply@anthropic.com>
### What this PR does / why we need it?
This PR fixes the logger initialization in patches so that the log info
can be displayed as expected.
### Does this PR introduce _any_ user-facing change?
No.
- vLLM version: v0.17.0
- vLLM main:
4497431df6
---------
Signed-off-by: Angazenn <supperccell@163.com>
Co-authored-by: kunpengW-code <1289706727@qq.com>
Co-authored-by: linsheng1 <1950916997@qq.com>
### What this PR does / why we need it?
Currently, chunked prefill is forcibly enabled. DeepSeek V3.1 W8A8C8
supports only the PD separation scenario. C8 refers to quantizing the KV
cache to int8, which aims to reduce the GPU memory usage of the KV cache
and improve the inference throughput.
Constraints:
1. Only the PD separation mode can be used and
MooncakeLayerwiseConnector can be used to run the model.
2. Currently, only the activation value supports dynamic quantization,
and the KV cache supports static quantization. C8 quantization with MTP
is not supported. You can use ModelSlim for quantization. The
quantization procedure is as follows:
pip install transformers==4.48.2
git clone https://gitcode.com/Ascend/msmodelslim.git
cd msmodelslim
bash install.sh
cd example/DeepSeek/
python3 quant_deepseek_w8a8.py --model_path <path/weight> --save_path
<path/quant_weight>
--anti_dataset../common/deepseek_anti_prompt_50_v3_1.json
--calib_dataset../common/deepseek_calib_prompt_50_v3_1.json --rot
--trust_remote_code True --fa_quant --dynamic --anti_method m6
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: pichangping <1337510399@qq.com>
Signed-off-by: Wang Kunpeng <1289706727@qq.com>
Co-authored-by: Wang Kunpeng <1289706727@qq.com>
### What this PR does / why we need it?
This PR restores #7029, which adds W8A8C8 support for dsv3.2/glm5 using
the `lightning_indexer_quant` ops in the pd-mix stage.
The original PR was reverted by #7288 because the patch did not work
with the recompute scheduler.
This PR also fixes the patching issue so that it works correctly with
the recompute scheduler.
### Does this PR introduce _any_ user-facing change?
Yes. To enable LI C8, users need to set the `enable_sparse_c8` option to
`"true"` in `additional_config`.
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: rjg-lyh <1318825571@qq.com>
### What this PR does / why we need it?
This reverts commit 7ed9e9de69, which
introduces an issue that the patch doesn't work with recompute scheduler
enabled.
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
To support prefix cache for Qwen3.5/Next in vLLM-Ascend, this PR mainly
follows the design in
[#30877](https://github.com/vllm-project/vllm/pull/30877) and inherits
changes to functions which are overridden in vLLM-Ascend.
Note:
1. `--mamba-cache-mode align` && PD disaggregation is still not
supported yet in vLLM v0.17.0(see
https://github.com/vllm-project/vllm/blob/main/vllm/v1/core/sched/scheduler.py#L295).
2. The current implementation of hybrid kv cache might result in a very
large block_size when scheduling. For example, if we run Qwen3.5-35B-A3B
with `-tp 2`, the block_size is adjusted to 2048, which means that any
prefix shorter than 2048 will never be cached. Although this behavior is
consistent with vLLM, it still needs improvements in the future.
3. `--mamba-cache-mode align` requires to copy mamba states during
forward steps. vLLM uses a triton kernel to implement it. However, the
original version run into some bugs on Ascend hardwares. Thus we patch a
new triton kernel to avoid this bug.
### Does this PR introduce _any_ user-facing change?
To use mamba prefix cache, set `--enable-prefix-caching` and
`--mamba-cache-mode align`. Note that the mamba state copy function(see
[do_mamba_copy_block](https://github.com/vllm-project/vllm/blob/main/vllm/v1/worker/mamba_utils.py#L132))
does not provide a torch native version, thus it might have trouble if
users can't use triton.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: Angazenn <supperccell@163.com>
### What this PR does / why we need it?
Drop 0.16.0 support in main
- Fix eagle proposer break introduced by
https://github.com/vllm-project/vllm/pull/34552. Mainly change to use
the draft attention group to initialize the attention metadata builder.
- Fix the `ModelRunner` has no attribute `cudagraph_capture_sizes`
error, which is a bug in vLLM v0.17.0, and fixed by a later pr
https://github.com/vllm-project/vllm/pull/30515
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
This PR supports W8A8C8 in dsv3.2/glm5 with lightning_indexer_quant ops
in pd-mix stage mainly.
Because the code for the current PD-disaggregated scenario is still
under refactoring and cleanup, this PR prioritizes ensuring the C8
functionality in the pd-mix scenario.
The next steps are planned in two parts:
① Once the optimized scatter operator is updated, we will replace the
original operator to improve the performance of storing k_scale.
② Once the code logic for the PD-disaggregated scenario becomes stable,
we will carry out more comprehensive validation and make appropriate
adaptations.
③ Because enabling C8 currently introduces several new operators whose
performance still needs improvement, performance may regress in some
scenarios. Therefore, only after all the operators are fully ready can
we ensure that this feature does not cause any performance degradation.
At that point, we will enable this feature by default and remove the
switch in `additional_config`.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
CI passed with new added/existing test.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: rjg-lyh <1318825571@qq.com>
### What this PR does / why we need it?
This PR aims to support aclgraph for model runner v2, please see RFC
#5208. The PR contains these modifications:
- adapt to newest commit of vllm main branch.
- supply a unified interface of extra forward context for both model
runner v1 and model runner v2.
- implement graph mode for main model.
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
### What this PR does / why we need it?
When GLM5 target model uses rotary quant, the final hidden states passes
to MTP need to do an extra rotary.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: Wangbingjie <wangbj1207@126.com>
Signed-off-by: wangbj127 <256472688+wangbj127@users.noreply.github.com>
### What this PR does / why we need it?
Fixed the error of speculative decoding in FULL mode when `num_spec + 1`
not in `cudagraph_capture_sizes`.
Now, we can run speculative decoding in FULL mode, but with drafter as
eager.
It depends on https://github.com/vllm-project/vllm-ascend/pull/7144 .
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
Test code is shown as below:
```python
prompts = [
"1.Who are you?",
"2. Who are you?",
]
sampling_params = SamplingParams(temperature=0.0, top_p=0.95, top_k=40, max_tokens=200)
llm = LLM(
model="/home/some-model/Meta-Llama-3.1-8B-Instruct",
tensor_parallel_size=1,
max_num_seqs=32,
# enforce_eager=True,
disable_log_stats=False,
distributed_executor_backend="mp",
gpu_memory_utilization=0.7,
async_scheduling=True,
speculative_config={
"enforce_eager": True,
"model": "/home/some-model/EAGLE3-LLaMA3.1-Instruct-8B",
"disable_padded_drafter_batch": False,
"method": "eagle3",
"num_speculative_tokens": 2,
},
compilation_config={
"cudagraph_mode": "FULL",
"cudagraph_num_of_warmups": 1,
},
max_model_len=4096,
enable_prefix_caching=False,
)
outputs = llm.generate(prompts, sampling_params)
```
The result before:
```text
File "/vllm-workspace/vllm/vllm/v1/cudagraph_dispatcher.py", line 140, in _create_padded_batch_descriptor
assert num_tokens_padded % uniform_decode_query_len == 0
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError
```
The result after:
```text
--------------------------------------------------
total_num_output_tokens: 400
num_drafts: 249
num_draft_tokens: 498
num_accepted_tokens: 149
mean acceptance length: 1.60
--------------------------------------------------
acceptance at token 0: 0.43
acceptance at token 1: 0.17
```
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
Signed-off-by: drslark <slarksblood@qq.com>
### What this PR does / why we need it?
Initial version to support minimax-m2.5 on vllm-ascend.
This commit coverting original fp8 weight to a quantilized bf16 to
support Minimax-m2.5 on NPU.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
### Test Report
Self tested precision summary, where the official precision score of
AIME2025 is 86.3
<img width="426" height="84" alt="image"
src="https://github.com/user-attachments/assets/a3ce2452-92fa-4713-962e-862248e0b61a"
/>
---------
Signed-off-by: limuyuan <limuyuan3@huawei.com>
Signed-off-by: SparrowMu <52023119+SparrowMu@users.noreply.github.com>
Co-authored-by: limuyuan <limuyuan3@huawei.com>
### What this PR does / why we need it?
Mooncake Layerwise Connector supports hybrid attention manager with
multiple kvcache groups.
### Does this PR introduce _any_ user-facing change?
Yes.
### How was this patch tested?
By CI.
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
---------
Signed-off-by: nwpu-zxr <zhouxuerong2@huawei.com>
### What this PR does / why we need it?
The ops `torch_npu.npu_recurrent_gated_delta_rule` currently does not
support `ssm_state` inputs in float32 format,
we temporarily retain the _forward_core implementation with triton for
Qwen3_5
---------
Signed-off-by: pppeng <zepengliu912@qq.com>
Signed-off-by: pppeng <60355449+ppppeng@users.noreply.github.com>
### What this PR does / why we need it?
This PR fixes a bug in the `_merge_multimodal_embeddings` function where
the parameter order was incorrect. The `multimodal_embeddings` and
`is_multimodal` parameters were swapped, which would lead to runtime
errors when the function is called with positional arguments.
This change corrects the function signature to align with its expected
usage, ensuring that multimodal embeddings are correctly merged.
### Does this PR introduce _any_ user-facing change?
No. This is a bug fix for an internal utility function and has no
user-facing impact.
### How was this patch tested?
The correctness of this fix is validated by existing tests for
multimodal functionality. With the incorrect function signature, these
tests would fail due to argument type mismatches. CI passing confirms
the fix is effective.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
Signed-off-by: tanhaoan333 <tanhaoan@huawei.com>
### What this PR does / why we need it?
Supports contiguous tensor hybrid-attn kv-cache on fullattn-mamba hybrid
model, such as Qwen3Next and Qwen3.5.
Due to the restrictions of Ascend operators, all KV tensors, conv
tensors, and SSM tensors must be contiguous. Therefore, this PR uses the
following solution to generate the KV cache:
tensor1: [(kv_padding), conv , ...]
tensor2: [k , ssm , ...]
tensor3: [v , (mamba_padding), ...]
Under this scheme, although some waste may occur, the tensors of all
caches are guaranteed to be contiguous.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
By CI.
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
---------
Signed-off-by: nwpu-zxr <zhouxuerong2@huawei.com>
### What this PR does / why we need it?
Change recurrent_gated_delta_rule ops from triton to ascend C version
for better performance.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: v0.15.0
- vLLM main:
9562912cea
---------
Signed-off-by: SunnyLee219 <3294305115@qq.com>
### What this PR does / why we need it?
If some `eagle3` model without embed_tokens works with `quarot` target
model, the acceptence rate will drop.
We solve it in this PR.
The relative vllm pr is https://github.com/vllm-project/vllm/pull/36225.
- vLLM main:
4034c3d32e
Signed-off-by: drslark <slarksblood@qq.com>
