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>
<!-- 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?
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>
### 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>
### 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?
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>
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>
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?
Fixed the bug of incorrect reshape usage.
For example:
ori_tensor: [[1, 2, 3], [4, 5, 6]]
after reshape:
[[1, 2], [3, 4], [5, 6]]
after permute:
[[1, 4], [2, 5], [3, 6]]
Now, we will directly use squeeze for a more intuitive understanding.
pr for main:
#7887
### Does this PR introduce _any_ user-facing change?
The actual peak-to-average ratio has successfully decreased.
Signed-off-by: shenchuxiaofugui <1311027364@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?
Before when we do put for KV Pool, we find the first non-existing key
and put all the blocks starting from that index; however, if the prefix
cache blocks is from another request, and some of the blocks are evicted
due to LRU, we will be putting blocks that still exist in the pool, and
causing MooncakeStore printing unnecessary logs in master service.
What this PR does:
Now we lookup all the keys and only put the ones that are missing.
Fix lookup_scheduler in pool_worker so it handles GQA correctly.
Fixes a few existing typos
Add UT, written by codex
<!--
- 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>
## 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?
cherry-pick from https://github.com/vllm-project/vllm-ascend/pull/7736
**Error information**
When the quantized weights in CompressedTensors format of the kimi-k2
model are used, the following error is reported:
`AttributeError: 'AscendCompressedTensorsConfig' obiect has no attribute
'enabling_fa_quant'`
**Error Cause**
Currently, FA3 quantization supports only the weights of modelslim
quantization. The added methods are not defined in
AscendCompressedTensorsConfig.
**Solution**
Before invoking related methods, check whether the FA3 feature is
enabled.
Additionally, the unused `get_scaled_act_names` method and its
corresponding unit test have been removed.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Existing unit tests were updated by removing a deprecated test case, and
the refactored logic was reviewed for correctness.
Signed-off-by: Wang Kunpeng <1289706727@qq.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?
Fixed the issue where the DCP overlaps the MTP scenario in the ds3.2
scenario.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
cherry-pick from: https://github.com/vllm-project/vllm-ascend/pull/7617
Signed-off-by: weiguihua2 <weiguihua2@huawei.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>
### What this PR does / why we need it?
This PR lets NPU platform provide its own default
`max_cudagraph_capture_size` via
`NPUPlatform.apply_config_platform_defaults()`.
Previously, when cudagraph sizing was left unset, Ascend inherited
vLLM's upstream default heuristic in `_set_cudagraph_sizes()`, which
uses `max_num_seqs * decode_query_len * 2`. This PR changes Ascend's
default to `min(max_num_seqs * decode_query_len, 512)` while keeping the
rest of vLLM's cudagraph sizing logic unchanged.
### Does this PR introduce _any_ user-facing change?
Yes, but only for Ascend when users do not explicitly configure
cudagraph sizing.
If `max_cudagraph_capture_size` and `cudagraph_capture_sizes` are both
unset, we now uses `max_num_seqs * decode_query_len` (capped at `512`)
instead of the upstream `* 2` default. Explicit user settings are
unchanged.
### How was this patch tested?
Add unit tests to cover:
- default max injection via `apply_config_platform_defaults()`
- explicit `max_cudagraph_capture_size` is preserved
- explicit `cudagraph_capture_sizes` are preserved
- Ascend default max no longer uses the upstream `* 2`
- late `_set_cudagraph_sizes()` recomputation reuses the current max
input
- vLLM version: v0.18.0
- vLLM main:
ed359c497a
---------
Signed-off-by: linfeng-yuan <1102311262@qq.com>
### What this PR does / why we need it?
RFC https://github.com/vllm-project/vllm-ascend/issues/7394
Add a PyTorch implementation of the chunk gated delta rule on 310P.
### Does this PR introduce _any_ user-facing change?
NO
### How was this patch tested?
UT
---------
Signed-off-by: Tflowers-0129 <2906339855@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?
This PR adds missing arguments in `AscendRotaryEmbedding`,
`AscendYarnRotaryEmbedding` to conform with vLLM. Besides, corresponding
ut is introduced.
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: Angazenn <supperccell@163.com>
### What this PR does / why we need it?
RFC #7394
310P cannot use the fused `rmsnormgated` operator and must fall back to
the native implementation.
### Does this PR introduce _any_ user-facing change?
NO
### How was this patch tested?
ut
- vLLM version: v0.17.0
- vLLM main:
4497431df6
---------
Signed-off-by: Tflowers-0129 <2906339855@qq.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?
Because the new A5 MMEncoder operator was merged, the 310P can no longer
run any VL models. This PR fixes that issue. details at #7046
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
e2e
- vLLM version: v0.17.0
- vLLM main:
8b6325758c
---------
Signed-off-by: Tflowers-0129 <2906339855@qq.com>
### What this PR does / why we need it?
Follow https://github.com/vllm-project/vllm/pull/37425,
https://github.com/vllm-project/vllm-omni/pull/1982
Copied from them:
Notice that `hasattr(self.model, "flush_pending_metadata")` cost 6ms per
decode step when profiling Qwen3 Omni.
The original `CUDAGraphWrapper.__getattr__` raises:
```python
raise AttributeError(f"... cudagraph wrapper: {self.runnable}")
```
When hasattr() is called for a non-existent attribute, Python internally
calls __getattr__ which constructs this AttributeError. The
{self.runnable} triggers `__repr__()` on the underlying model (e.g.,
`Qwen3OmniMoeForConditionalGeneration`), which recursivelytraverses the
entire nn.Module tree to generate an 18,000+ character string. This
takes ~6-7ms per call.
Since `hasattr(self.model, "flush_pending_metadata") ` is called every
decode step in the Talker forward path, this adds ~6ms overhead per
step, severely impacting audio inter-chunk latency (ICL).
```Python
hasattr(self.model, "flush_pending_metadata")
→ getattr(self.model, "flush_pending_metadata")
→ not found in CUDAGraphWrapper.__dict__
→ not found in the CUDAGraphWrapper class hierarchy
→ triggers CUDAGraphWrapper.__getattr__("flush_pending_metadata")
→ hasattr(self.runnable, "flush_pending_metadata") # runnable also doesn't have it
→ executes raise AttributeError(f"... {self.runnable}")
→ Python needs to construct the exception object
→ the f-string triggers self.runnable.__repr__()
→ Qwen3OmniMoeForConditionalGeneration.__repr__()
→ recursively traverses the entire nn.Module tree
→ generates a 18,000+ character string
→ takes ~6 ms
→ AttributeError object is created
→ hasattr catches the AttributeError and returns False
→ the 18,000-character string is immediately discarded (no one ever sees it)
```
### Does this PR introduce _any_ user-facing change?
NO.
### How was this patch tested?
See https://github.com/vllm-project/vllm-omni/pull/1982
- vLLM version: v0.17.0
- vLLM main:
4497431df6
---------
Signed-off-by: gcanlin <canlinguosdu@gmail.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
This PR fixes a startup regression for Ascend hybrid attention + mamba
models after upgrading to vLLM `0.18.0`.
However, after the vLLM `0.18.0` upgrade, worker initialization still
calls the generic platform hook:
- `current_platform.update_block_size_for_backend(vllm_config)`
### How this PR fixes it
This PR keeps the fix strictly inside `vllm-ascend`.
It adds an Ascend override for
`NPUPlatform.update_block_size_for_backend()`:
- for hybrid models, do not run the generic upstream block-size fallback
- preserve the block size that was already computed by the hybrid
model-specific config logic
- for non-hybrid models, keep the original upstream behavior unchanged
- vLLM version: v0.18.0
- vLLM main:
8b6325758c
---------
Signed-off-by: maoxx241 <maomaoyu870@gmail.com>
Signed-off-by: Mengqing Cao <cmq0113@163.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
### What this PR does / why we need it?
This PR aims to fix padding logic in eagle proposer for kimi25. Main
changes involve:
1. modify the way to obtain draft model attention builder and backend
2. add block table padding & related tensor slicing in common metadata
when `draft_step>1` for solving fia verifying error
3. replace block table in `update_graph_params` for solving fia
verifying error
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
Signed-off-by: Zetong Li <slippersss@126.com>
### What this PR does / why we need it?
This PR enables separate attention backend configuration for target and
draft models in speculative decoding, decoupling the previously bound
attention backend settings between the two models.
It solves the compatibility issue where some draft models do not support
the attention backend used by the target model, and allows users to
select the optimal attention backend for each model individually to
maximize inference performance. The change is fully backward compatible.
---------
Signed-off-by: SidaoY <1024863041@qq.com>
### What this PR does / why we need it?
Refactor `vllm_ascend/ops/fused_moe` to replace scattered MoE business
`**kwargs` with typed request objects and explicit stage boundaries.
- Prepare, dispatch, MLP, and quant stages now have clearer ownership.
- Main MoE path no longer depends on business `kwargs.get(...)` lookups.
- Comm and dispatcher interfaces are request-only on the main path.
- UTs can assert stage-level fields directly instead of inferring
behavior indirectly.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
CI passed.
---------
Signed-off-by: linfeng-yuan <1102311262@qq.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?
**Refactor: Replace npu_ring_mla with FIA in MLA prefill**
This PR refactors the MLA (Multi-Layer Attention) prefill implementation
by replacing `npu_ring_mla` with `npu_fused_infer_attention_score` (FIA)
operator, unifying the attention backend with the standard attention
implementation.
**Key changes:**
1. **Core prefill refactoring (`mla_v1.py`)**
- Replace `npu_ring_mla` with `npu_fused_infer_attention_score` in
`_forward_prefill` and `_compute_prefill_context`
- Use TND layout with `softmax_lse_flag=True` for prefill attention
- Use `npu_attention_update` to merge multiple chunk outputs with LSE
(Log-Sum-Exp)
- Change `attn_mask` from `get_final_mla_mask()` to
`get_splitfuse_attn_mask()` for FIA compatibility
2. **Data type handling**
- Add automatic float16 → bfloat16 conversion (FIA with TND layout only
supports bfloat16)
- Convert output back to original dtype after FIA computation
3. **Metadata optimization**
- Pre-calculate `actual_seq_lengths_q` in `AscendMLAPrefillMetadata`
- Pre-calculate `chunk_actual_seq_lengths_kv_list` in
`ChunkedContextMetadata`
- Move `torch.cumsum` operations from forward pass to metadata building
phase
4. **CP compatibility (`mla_cp.py`)**
- Add `_ring_mla_mask_builder` to get `npu_ring_mla`-compatible masks
for Context Parallel scenarios
- Add `chunk_actual_seq_lengths_kv_list` field to
`CPChunkedContextMetadata`
**Why we need it:**
- **Backend unification**: Aligns MLA prefill with standard attention
implementation (`attention_v1.py`)
- **Better chunked context support**: FIA + `npu_attention_update`
provides native LSE-based output merging
- **Future compatibility**: Prepares for eventual `npu_ring_mla` removal
across the codebase
### Does this PR introduce _any_ user-facing change?
**No.** This is a pure refactoring with no functional changes - same
behavior, unified backend.
---
- Related issue: #5463 (item 7)
- vLLM version: v0.14.1
Signed-off-by: lico67373 <918688502@qq.com>
### What this PR does / why we need it?
Reapply the auto-detect quantization format feature (originally in
#6645, reverted in #6873) and extend it to support remote model
identifiers (e.g., `org/model-name`).
Changes:
- Reapply auto-detection of quantization method from model files
(`quant_model_description.json` for ModelSlim, `config.json` for
compressed-tensors)
- Add `get_model_file()` utility to handle file retrieval from both
local paths and remote repos (HuggingFace Hub / ModelScope)
- Update `detect_quantization_method()` to accept remote repo IDs with
optional `revision` parameter
- Update `maybe_update_config()` to work with remote model identifiers
- Add platform-level `auto_detect_quantization` support
- Add unit tests and e2e tests for both local and remote model ID
scenarios
Closes#6836
### Does this PR introduce _any_ user-facing change?
Yes. When `--quantization` is not explicitly specified, vllm-ascend will
now automatically detect the quantization format from the model files
for both local directories and remote model IDs.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: SlightwindSec <slightwindsec@gmail.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 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?
New Quantization Method: Introduced support for the W8A8SC static linear
quantization scheme specifically for 310P hardware, enabling more
efficient model compression.
Refactored the save_sharded_state_310.py to avoid multi-process issue.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
W8A8SC quant E2E test.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: pu-zhe <zpuaa@outlook.com>
### What this PR does / why we need it?
Support chunked prefill for Qwen3Next with PCP&DCP
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
---------
Signed-off-by: QiuChunshuo <qiuchunshuo@huawei.com>
### What this PR does / why we need it?
This pull request addresses a bug related to the fused mc2 functionality
within the EPLB (Expert Parallelism Load Balancing) system, specifically
impacting quantization and MoE communication.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.15.0
- vLLM main:
83b47f67b1
Signed-off-by: Spicy-Stick <873805887@qq.com>
Signed-off-by: root <root@localhost.localdomain>
### What this PR does / why we need it?
**NOTE: This PR is re-pull of #7016 since ci mistakenly marked
unfinished pr as having passed.**
This PR aims to delete mtp_proposer. By fixing a bug in both dsv32 and
glm5, now it should be ok to remove mtp_proposer. The bug is actually
about unnecessary slicing of `slot_mapping`.
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
by ci
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: Zetong Li <slippersss@126.com>
### What this PR does / why we need it?
This PR aims to delete mtp_proposer. By fixing a bug in both dsv32 and
glm5, now it should be ok to remove mtp_proposer. The bug is actually
about unnecessary slicing of `slot_mapping`.
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
by ci
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
---------
Signed-off-by: Zetong Li <slippersss@126.com>
## What this PR does / why we need it?
Implements [RFC
#6954](https://github.com/vllm-project/vllm-ascend/issues/6954):
NPUWorker Profiler profile_prefix full adaptation for API parity with
upstream vLLM.
### Changes
- **Lazy profiler init**: Defer profiler creation until first
`profile(is_start=True)` call
- **profile_prefix param**: Add `profile_prefix` to `profile()`; compute
`trace_name` from prefix + `get_worker_rank_suffix()`
- **Refactor `_init_profiler` → `_create_profiler(trace_name)`**: Pass
`worker_name` to `tensorboard_trace_handler` for unique trace files per
worker
- Unique trace files per worker; no collision in multi-worker setups
### Testing
- Unit tests updated/added in `tests/ut/worker/test_worker_v1.py`
- `pytest tests/ut/worker/test_worker_v1.py::TestNPUWorker` passed
## Does this PR introduce _any_ user-facing change?
Yes. Trace file naming may differ (more descriptive with worker rank
suffix). `profile(is_start=True, profile_prefix="warmup")` now
supported.
## How was this patch tested?
- Unit tests:`pytest tests/ut/worker/test_worker_v1.py::TestNPUWorker`
- Manual: vLLM serve with profiler config, start/stop profile, verified
trace files
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
---------
Signed-off-by: realliujiaxu <realliujiaxu@163.com>
This pull request refactors the speculative decoding proposer interface
to align with upstream vLLM, removing the local `Proposer` interface and
renaming methods to `propose`.
This is the first step. In the future we should remove the class
register and just add few Ascend specified method once the arch in vLLM
is ready.
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
This PR add docs of batch invariance and make some extra operators
according to validation result.
please see https://github.com/vllm-project/vllm-ascend/issues/5487 to
track progress.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
---------
Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
