### What this PR does / why we need it?
Add new function to mooncake layerwise connector, including:
1. supports sparse attention, for DeepSeek-V3.2
2. Distribute transfer tasks to redundant kv_head cards
This PR is related to [[RFC]: CDCP Scheduling for Disaggregated
Prefilling with KV Cache Layerwise Push
Support](https://github.com/vllm-project/vllm-ascend/issues/4842)
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
By CI.
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: nwpu-zxr <zhouxuerong2@huawei.com>
Signed-off-by: liziyu <liziyu16@huawei.com>
Co-authored-by: liziyu <liziyu16@huawei.com>
### What this PR does / why we need it?
The rotary algorithm in deepseek indexer should be neox-style instead of
gptj style. PR #4413 fix this accuracy bug with new triton kernel. This
PR fixes original pytorch version.
### Does this PR introduce _any_ user-facing change?
None
### How was this patch tested?
CI passed with existing test.
- vLLM version: 86e178f7c4d8c3b0eaf3c8e3f810a83f63b90e24
- vLLM main:
86e178f7c4
Signed-off-by: whx-sjtu <2952154980@qq.com>
### What this PR does / why we need it?
mlapo in deepseek is a huge performance improvement in decode, this pr
support pcp & dcp with mlapo
### Does this PR introduce _any_ user-facing change?
NO
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: zhenwenqi2024 <zhenwenqi_2022@qq.com>
### What this PR does / why we need it?
This addresses the issue brought up by #5356 and #4963, and we believe
the unnecessary conditions are the root cause.
Change the unpad trigger to be driven by actual size mismatches
(num_reqs vs base_num_reqs or scheduled vs input token counts) rather
than specific speculative-method flags. Then remove brittle workarounds
that forced request counts and sliced query start locations.
This prevents incorrect indexing and length mismatches during
speculative decoding and makes metadata unpadding more robust across
scheduling modes.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
Tested by existing cases.
- vLLM version: v0.13.0
- vLLM main:
8be6432bda
---------
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
#5230 this PR introduced a problem when both mtp and full_decode_only
are enabled for the DSV32 model, the operators cannot be compiled into
the graph. This PR fixes that issue.
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
Signed-off-by: cookieyyds <126683903+cookieyyds@users.noreply.github.com>
### What this PR does / why we need it?
- Delete the environment variable
`VLLM_ASCEND_ENABLE_FLASHCOMM2_OSHARED`
- Introduce layer_sharding as a configurable feature in
additional_config
- Revise the term "shared weight" to "shard weight."
Configuration : The feature is opt-in via the additional_config
argument:
```
--additional-config '{
"layer_sharding": ["o_proj", "q_b_proj"]
}'
```
This is orthogonal to standard tensor parallelism and weight replication
strategies. It is treated as a separate, explicit feature.It can be used
in any scenario, combined with the
flashcomm2https://github.com/vllm-project/vllm-ascend/pull/3232 feature
or the ShardedCP #4702 feature, to achieve significant performance.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: zzhx1 <zzh_201018@outlook.com>
Signed-off-by: zzhxx <zhangzihang23@mails.ucas.ac.cn>
Signed-off-by: chenxiao <Jaychou1620@Gmail.com>
Co-authored-by: clrs97 <524936896@qq.com>
Co-authored-by: Levi-JQ <yujinqi2@huawei.com>
Co-authored-by: chenxiao <Jaychou1620@Gmail.com>
## What this PR does / why we need it?
This PR fixes the `AttentionMaskBuilder` singleton initialization issue
introduced in PR #4779 and removes the unused `pcp_prefill_mask` field.
### Background
After PR #4779 made `AttentionMaskBuilder` a singleton with `@singleton`
decorator, the class constructor now requires a `device` parameter.
However, two initialization sites were still using the old parameterless
constructor, causing failures.
### Changes
1. **Fix singleton initialization**
- Fixed `AttentionMaskBuilder()` → `AttentionMaskBuilder(self.device)`
in `AscendMLAMetadataBuilder.__init__()`
- Fixed `AttentionMaskBuilder()` → `AttentionMaskBuilder(self.device)`
in `AscendAttentionMetadataBuilder.__init__()`
2. **Remove unused field**
- Removed `pcp_prefill_mask` field from
`AscendPrefillContextParallelMetadata` (never used in codebase)
- Updated related test assertions
### Related
- Issue #5463
- PR #4779 (Unify all mask generation methods)
- PR #5389 (Make AttentionMaskBuilder singleton)
## Does this PR introduce _any_ user-facing change?
No. This is an internal refactoring.
## How was this patch tested?
- ✅ Local testing: No linter errors
- ✅ Unit tests for attention modules verified
- ⏳ CI pipeline
Signed-off-by: lico67373 <918688502@qq.com>
Co-authored-by: weijinqian0 <1184188277@qq.com>
### What this PR does / why we need it?
When using the swa parameter in fia, headDim does not currently support
256, and when gemma3's headDim is equal to 256, an error will occur.
Therefore, code rollback is required, and it will be incorporated after
cann supports it.
### Does this PR introduce _any_ user-facing change?
Remove swa parameter of fia.
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
7157596103
---------
Signed-off-by: nsdie <yeyifan@huawei.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
### What this PR does / why we need it?
- Problem: In MLA+MLAPO, KV-consumer deployments keep
fused_qkv_a_proj/q_proj weights and quant params even though MLAPO uses
the prepacked buffers, increasing memory footprint on decode nodes.
- Fix: Conditionally drop those tensors only when
`kv_transfer_config.is_kv_consumer` to reclaim memory (consistent with
the SFA behavior #4774 ).
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: Chen Chen <0109chenchen@gmail.com>
### What this PR does / why we need it?
Since the _npu_ring_mla operator deteriorates in long-sequencescenarios,
the long sequence is split into shorter sequences for input to improve
performance.
- vLLM version: v0.13.0
- vLLM main:
5326c89803
---------
Signed-off-by: pichangping <1337510399@qq.com>
### What this PR does / why we need it?
This PR adds multi-stream for GQA to enable computation-communication
overlap. For chunked prefill, we reduce TTFT by approximately 4%.
### Does this PR introduce _any_ user-facing change?
No
- vLLM version: release/v0.13.0
- vLLM main:
bc0a5a0c08
---------
Signed-off-by: QiuChunshuo <qiuchunshuo@huawei.com>
### What this PR does / why we need it?
Supported to use full-graph with Qwen3-Next-MTP.
In detail, we adatpted `AscendAttentionState.ChunkedPrefill` in main
model, and also adapted `AscendAttentionState.ChunkedPrefill` in mtp
model.
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
We changed the test of Qwen3-Next-MTP in
`tests/e2e/multicard/test_qwen3_next.py` to make it a test of
`FULL_DECODE_ONLY`. Then run `pytest -s
tests/e2e/multicard/test_qwen3_next.py::test_qwen3_next_distributed_mp_eager_mtp_similarity_tp4`.
And this test passed.
```text
.
================================================================================================================================= warnings summary =================================================================================================================================
<frozen importlib._bootstrap>:241
<frozen importlib._bootstrap>:241: DeprecationWarning: builtin type SwigPyPacked has no __module__ attribute
<frozen importlib._bootstrap>:241
<frozen importlib._bootstrap>:241: DeprecationWarning: builtin type SwigPyObject has no __module__ attribute
-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
==================================================================================================================== 1 passed, 2 warnings in 271.89s (0:04:31) =====================================================================================================================
```
- vLLM version: v0.13.0
- vLLM main:
5326c89803
Signed-off-by: drslark <slarksblood@qq.com>
### What this PR does / why we need it?
In the current process of implementing attention updates, the FIA
operator shares a single workspace among different layers within the
same computation graph. To enable memory reuse, we adopt the
weak_ref_tensor mechanism. However, this approach may lead to precision
anomalies in certain scenarios. To address this issue, different layers
in the same computation graph are assigned independent workspaces.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
45c1ca1ca1
Signed-off-by: WithHades <244036962@qq.com>
### What this PR does / why we need it?
Improve the performance of Layerwise Connector, mainly includes the
following points:
1. Use event synchronize to replace stream synchronize.
2. Access metaserver when scheduling.
3. Transfer kvcache each Chunk prefill segmentation.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
By CI.
- vLLM version: release/v0.13.0
- vLLM main:
5fbfa8d9ef
---------
Signed-off-by: nwpu-zxr <zhouxuerong2@huawei.com>
Signed-off-by: liziyu <liziyu16@huawei.com>
Signed-off-by: wangxiaoteng <wangxiaoteng@huawei.com>
Co-authored-by: liziyu <liziyu16@huawei.com>
Co-authored-by: wangxiaoteng <wangxiaoteng@huawei.com>
By converting the KV cache from ND to NZ format when the decode node
receives it, this PR ensures that the KV NZ feature works correctly
during the decoding phase in disagg-prefill scenario.
- vLLM version: v0.11.0
- vLLM main:
83f478bb19
---------
Signed-off-by: Jade Zheng <zheng.shoujian@outlook.com>
Co-authored-by: ghphotoframe <854746559@qq.com>
Co-authored-by: alex101-ops <alex1015718386@gmail.com>
### What this PR does / why we need it?
Refactor the `capture_model` method in model_runner to directly reuse
the method from vLLM.
Currently, most of the logic in the capture_model method is similar to
that in the vllm code. Directly using the vllm method can reduce the
maintenance cost of the vllm-ascend code. Modify as follows:
1、refactor capture_model function, directly inheriting community methods
2、refactor initialize_aclgraph_capture function, move to
initialize_attn_backend
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
Signed-off-by: weiguihua2 <weiguihua2@huawei.com>
Co-authored-by: weijinqian0 <1184188277@qq.com>
### What this PR does / why we need it?
#5051 only implement a basic framework for model runner v2, but there
are still some bugs for e2e functionality, this PR aim to enable basic
functionality.
model runner v2 plans:
https://github.com/vllm-project/vllm-ascend/issues/5208
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
Enable fia to support sliding window function and adapt to the Gemma3
model.
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: nsdie <yeyifan@huawei.com>
### What this PR does / why we need it?
We support to use full graph with eagle.
Change list:
1. Distinguish between processing graph_params and draft_graph_params in
attention_v1.
2. Adapt the full-graph mode in eagle_proposer, include:
1). If use full graph, make Fullgraph Wrapper when load model.
2). Build a new meatadata, set running mode in FULL and mark attention
update in dummy_run when in Fullgraph mode.
3). Fixed and fill any attn_metadata, such as
attn_metadata.slot_mapping.
4). Add a descriptor.
5). Set running mode and triggered update metadata.
3. Trans is_mtp_model to is_draft_model, and add the update of
workspace.
NOTE:
When set async_scheduling=True, the draft model will enforce execution
in eager mode.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: anon189Ty <Stari_Falcon@outlook.com>
Co-authored-by: Yizhou Liu <liu_yizhou@outlook.com>
Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
We'll release 0.13.0 soon. The main branch is freeze. Let's revert the
newest change and redo it once 0.13.0 is released
- vLLM version: release/v0.13.0
- vLLM main:
81786c8774
### What this PR does / why we need it?
Since the _npu_ring_mla operator deteriorates in long-sequencescenarios,
the long sequence is split into shorter sequences for input to improve
performance.
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: pichangping <1337510399@qq.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>
1. The `npu_fused_infer_attention_score` kernel supports specifying the
output layout. By selecting the appropriate layout, we can avoid the
transpose operation typically required after the attention.
2. The `transpose_batchmatmul` function allows us to control whether the
output tensor is transposed. If we configure `perm_y`, an additional
transpose after executing `v_up` becomes unnecessary.
- vLLM version: release/v0.13.0
- vLLM main:
254f6b9867
---------
Signed-off-by: Jade Zheng <zheng.shoujian@outlook.com>
### What this PR does / why we need it?
Currently, MHA models (eg: minicpm-2b, Baichuan-7b) will encounter
errors when running in piecewise graph mode, with error messages similar
to:
```
(E89999): When layout is TND and PA not enabled, keyT(8) and valueT(8) must be equal to the last element of actualSeqenceLengthKV(5)[FUNC:CheckInputShapeWhenLayoutIsTND][FILE:prompt_flash_attention_tiling.cpp][LINE:3618]
```
The error occurs because the qkv in the Prefill stage is also padded,
causing the shape to be inconsistent with actual_seq_lengths.
Add unpadding logic for kv.
- vLLM version: release/v0.13.0
- vLLM main:
254f6b9867
Signed-off-by: Wang Kunpeng <1289706727@qq.com>
### What this PR does / why we need it?
Revert [KV-Sharing] Support KV-Sharing feature in CLA models (#4138) as
it causes deepseek v3.2 hang error
- vLLM version: release/v0.13.0
- vLLM main:
5fbfa8d9ef
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
RFC: https://github.com/vllm-project/vllm-ascend/issues/4629
Reason：
The functions related to Cp differ significantly from those of normal
MLA-Attention, but the coupling is quite severe.
Steps：
1)Extract common code AscendMLAMetadataBuilder.build to 4 functions:
build_prefill_metadata, build_decode_metadata,build_cp_metadata,
build_chunked_metadata
todo：
1)refactor function _compute_prefill_context;
2)refactor function _mla_preprocess,_mla_decode_preprocess
3）Extract public data and processing functions from the attention_cp.py
and mla_cp.py files to the common_cp file.
vLLM version: 0.13.0rc3
vLLM main:
ad32e3e19c
- vLLM version: 0.13.0rc3
- vLLM main:
ad32e3e19c
---------
Signed-off-by: wujinyuan1 <wjy9595@qq.com>
Signed-off-by: wujinyuan1 <wujinyuan1@huawei.com>
Co-authored-by: wujinyuan1 <wjy9595@qq.com>
Co-authored-by: weijinqian0 <1184188277@qq.com>
### What this PR does / why we need it?
Support KV-Sharing feature in CLA (cross layer attention) models, which
sharing kv cache in some layers.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
Currently, Flashcomm1 and FULL_DECODE_ONLY are incompatible. When both
features are enabled, graph capture errors occur without clear error
messages.
After discussion, it has been determined that enabling FULL_DECODE_ONLY
with Flashcomm1 in mixed deployment scenarios provides almost no TPOT
benefit. Additionally, a reconstruction of the decode phase for
flashcomm1 is currently underway. Therefore, related adaptation work is
temporarily postponed and will be addressed after the decode phase
reconstruction plan is finalized.
For now, an assert will be added to provide clear error messages and
correct deployment recommendations.
### Does this PR introduce _any_ user-facing change?
NO
### How was this patch tested?
NO
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: QiuChunshuo <qiuchunshuo@huawei.com>
RFC: https://github.com/vllm-project/vllm-ascend/issues/4629
Reason:
The metadata data class contains an excessive number of variables. We
will inherit the metadata of the community and simultaneously remove
some variables that are no longer needed at present.
Todo:
1. remove attn_state partly.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: weijinqian_v1 <weijinqian@huawei.com>
Co-authored-by: weijinqian_v1 <weijinqian@huawei.com>
### What this PR does / why we need it?
Now `VLLM_ASCEND_ENABLE_NZ` will have three options:
0: disable nz;
1: only quant case enable nz;
2: enable nz as long as possible;
And `VLLM_ASCEND_ENABLE_NZ`=1 by default.
All cases are shown in the table below:
| | W4A4 | W4A8 | W8A8 | fp16/bf16 | fp32 |
|---|---|---|---|---|---|
| trans nz | can't support nz | trans nz by default | trans nz by
default | trans nz when VLLM_ASCEND_ENABLE_NZ is 2 | can't support nz |
| transpose | only support not transpose case | only support transpose
case | only support transpose case | linear: only support not transpose
case<br>gmm: only support transpose case | same to fp16/bf16 |
Some exceptional cases:
1. MLAPO op need to do some additional processing on the weights,
including trans nz. If use MLAPO op, some weight will be transformed to
nz forcely;
2. MLA/SFA's weight `W_UV` will be used by op
`torch.ops._C_ascend.batch_matmul_transpose`, and this op can't support
nz currently;
### Does this PR introduce _any_ user-facing change?
Now fp16/bf16 weight will not trans nz by default.
### How was this patch tested?
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: zzzzwwjj <1183291235@qq.com>
### What this PR does / why we need it?
Remove Pangu Related Code
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
e2e & ut
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: weichen <calvin_zhu0210@outlook.com>
### What this PR does / why we need it?
This PR fixes some incorrect `get_current_vllm_config` calling, which
creates empty vllm_config instead.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: Angazenn <supperccell@163.com>
### What this PR does / why we need it?
This PR fixes a bug in the `AscendMLAImpl._v_up_proj` method where the
optimized `batch_matmul_transpose` operator was not being utilized.
**Changes:**
- Modified `_v_up_proj` method to use
`torch.ops._C_ascend.batch_matmul_transpose` operator for FP16/BF16
dtypes when available
- Added fallback path using the original `torch.bmm` implementation for
other cases
- This avoids unnecessary transpose operations and improves performance
**Why needed:**
- The previous implementation only used `torch.bmm` with multiple
transpose operations, which is less efficient
- The Ascend backend provides an optimized `batch_matmul_transpose`
operator that can handle the computation more efficiently
- This fix improves inference performance for MLA (Multi-head Latent
Attention) models on Ascend NPU
### Does this PR introduce _any_ user-facing change?
No. This is a performance optimization that maintains the same
functionality and output. Users will experience faster inference for
MLA-based models, but no API or interface changes are introduced.
The changes maintain backward compatibility with the fallback path,
ensuring correct behavior when the operator is not available or for
unsupported dtypes.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: lico67373 <918688502@qq.com>
Co-authored-by: hwhaokun <haokun0405@163.com>
Co-authored-by: weijinqian0 <1184188277@qq.com>
RFC: https://github.com/vllm-project/vllm-ascend/issues/4629
Reason:
We distinguish the branches based on the applicable scenarios of
pagedAttention and fusedInferAttention, making the code more clear.
At the same time, it is convenient for the subsequent iterations of
sliding_window and sinks and removePA ops after FIA is ready.
Todo:
remove PA ops after FIA is ready
add slidingwindow and ops for gpt_oss
replace FIA with FIA_v2
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: weijinqian_v1 <weijinqian@huawei.com>
Co-authored-by: weijinqian_v1 <weijinqian@huawei.com>
### What this PR does / why we need it?
Rename `_910B` to `A2`;
Rename `_910_93` to `A3`;
Rename `_910_95` to `A5`;
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: zzzzwwjj <1183291235@qq.com>
### What this PR does / why we need it?
This PR add `qkv_rmsnorm_rope` operator and introduces a graph fusion
pass for `qknorm_rope` operations. The implementation includes a new
configuration flag, a pattern matching pass using
`torch._inductor.pattern_matcher`, and a custom Triton kernel for the
fused operation.
Co-authored-by: Angazenn
[supperccell@163.com](mailto:supperccell@163.com)
### Does this PR introduce _any_ user-facing change?
Yes, add new additional_config
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: wxsIcey <1790571317@qq.com>
### What this PR does / why we need it?
We refactored the eagle_proposer.py to adapt the framework of eagle.py
in vllm-v0.12.0, to support the logit of padded drafter batch and
async-scheduler.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: anon189Ty <Stari_Falcon@outlook.com>
Co-authored-by: drslark <slarksblood@qq.com>
### What this PR does / why we need it?
This PR adds back pa in scenarios of small batch sizes due to
performance consideration. Will remove pa once fia performs better than
pa in all scenarios.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
CI passed with existing test.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: whx-sjtu <2952154980@qq.com>
Co-authored-by: weijinqian0 <1184188277@qq.com>
