### What this PR does / why we need it?
refactor npu_modelrunner， we should be close to gpu_modelrunner
### Does this PR introduce _any_ user-facing change?
NO
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: zhenwenqi2024 <zhenwenqi_2022@qq.com>
Signed-off-by: zhenwenqi2024 <155598497+zhenwenqi2024@users.noreply.github.com>
### What this PR does / why we need it?
vllm-ascend support Ascend950 with Qwen dense model
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: wangyao <iwangyao@outlook.com>
Co-authored-by: weijinqian0 <1184188277@qq.com>
mindie_turbo is out of data for long time. This PR remove the related register method.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This PR clean up useless torchair logic in model runner. The moge doc is
only for torchair, it can be removed as well.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
VLLM_ASCEND_ENABLE_TOPK_TOPP_OPTIMIZATION is enabled by default for long
time. Let's remove it now.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
This PR eliminates the simplicit HD synchronization in sfa backend, and
_build_dummy_attn_metadata and dummy_run in mtp_proposer, significantly
improving dsv3.2 performance in low-latency scenarios.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Performance improvements are observed with E2E performance serving (P:
DP4TP8EP32 D: DP8TP4EP32) with `num_speculative_tokens=3`.
DSV3.2-W8A8-EXP:
TPOT: 41.67ms -> 23.36ms
ITL: 85.93ms -> 55.96ms
DSV3.2-W8A8 (relaesed in December):
TPOT: 18.11ms
ITL: 56.13ms
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: linfeng-yuan <1102311262@qq.com>
### What this PR does / why we need it?
We introduced the npugraph_ex backend through the vllm's adaptor
dispatch mechanism to accelerate aclgraph. This solution is based on
torch.compile and uses torchair to optimize the fx.graph. The
performance gains are mainly obtained from the static kernel. We
conducted tests on Qwen3-30B and achieved over 5% performance
optimization.
### Does this PR introduce _any_ user-facing change?
Yes, we add a new switch named"enable_npugraph_ex" in additional_config,
default is False.
We also add an example to show how to register custom replacement pass
### More information about this PR
This feature depends on the release of CANN and torch_npu in Q4.
We tested it on a package that has not been publicly released yet and
verified that the functionality works.
This feature is still experimental at the moment; setting the config
true will directly raise error.
Merging into the main branch initially involves some preliminary commits
to facilitate subsequent development and testing of the feature, as well
as to avoid submitting an excessively large PR at once.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: chencangtao <chencangtao@huawei.com>
Signed-off-by: ChenCangtao <50493711+ChenCangtao@users.noreply.github.com>
Co-authored-by: chencangtao <chencangtao@huawei.com>
Co-authored-by: panchao-hub <315134829@qq.com>
Co-authored-by: wbigat <wbigat@163.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
### What this PR does / why we need it?
The first commit support `FULL_DECODE_ONLY`:
- Update `AscendSFAMetadataBuilder` to use `num_input_tokens` for
slicing slots and positions, ensuring fixed tensor shapes.
- Implement padding logic for `query_start_loc` in `NPUModelRunner` to
support uniform decode in full graph mode, aligning with GPU runner
behavior.
- Adjust MLA cosine cache allocation to occur independently of graph
mode and switch to using device-resident sequence lengths for attention
metadata.
- Remove redundant slicing of hidden states and outputs in
`AscendSFAImpl` and optimize `sin`/`cos` cache updates.
The second commit take MTP into account:
- Update `AscendSFAMetadataBuilder` to use `num_input_tokens` for
slicing slots and positions, ensuring fixed tensor shapes.
- Implement padding logic for `query_start_loc` in `NPUModelRunner` to
support uniform decode in full graph mode, aligning with GPU runner
behavior.
- Adjust MLA cosine cache allocation to occur independently of graph
mode and switch to using device-resident sequence lengths for attention
metadata.
- Remove redundant slicing of hidden states and outputs in
`AscendSFAImpl` and optimize `sin`/`cos` cache updates.
And the rest of them are just bugfix.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
Test cases needed.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
The logprobs_tensor was not initialized before accessing its token_id
member, leading to a crash when tokenizer.decode() is called by passing
a negative token_id
### How was this patch tested?
Constructed an inference request with two prompts and set
SamplingParams(prompt_logprobs=<non-None value>) (e.g.,
prompt_logprobs=1).
After applying the fix (proper initialization of logprobs_tensor), the
same request completed successfully without errors, and the returned
logprobs matched expected values.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: jiangweixiang <jwx02384838@antgroup.com>
Co-authored-by: jiangweixiang <jwx02384838@antgroup.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
### What this PR does / why we need it?
To support pipeline parallel with PD disaggregation, this PR support PP
in mooncake connector and fix other bugs when enable pp with other
optimization params, including following changes:
- mooncake connector support pp in prefill, we do not support decode pp
currently
- fix bugs when enable both pp and flashcomm1
- optimize ascend-scheduler to support full batch in multiple pipeline
stages, original implementation would cause all pipeline stages
batch_size total summed to max_num_seq, which makes pipeline is not
full, this optimization can make all stages running with full batch_size
= max_num_seq, the same changes will contribute to vllm scheduler too.
### Does this PR introduce _any_ user-facing change?
add `pp_size` in mooncake connector kv_connector_extra_config
```
"kv_connector_extra_config": {
"use_ascend_direct": true,
"prefill": {
"dp_size": 1,
"tp_size": 4,
"pp_size": 4
},
"decode": {
"dp_size": 16,
"tp_size": 1
}
}
```
### How was this patch tested?
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: chenxiao <Jaychou1620@Gmail.com>
Signed-off-by: Kurumi5210 <Jaychou1620@Gmail.com>
Signed-off-by: Kurumi5210 <jaychou1620@gmail.com>
Signed-off-by: 秋刀鱼 <jaychou1620@Gmail.com>
Co-authored-by: chenxiao <Jaychou1620@Gmail.com>
Co-authored-by: zss <zss@qq.com>
Co-authored-by: zss <3265779424@qq.com>
### What this PR does / why we need it?
Adds W4A16 quantization method for the Kimi-K2-Thinking model and
updates relevant modules to support the new quantization method.
- Implements complete W4A16 quantization method including weight
packing/unpacking, per-group quantization parameter generation,
post-processing logic and MoE method application.
- Adds parameters `use_int4_w4a16`, `w1_offset` and `w2_offset`, adjusts
`with_quant` conditional logic to support W4A16 matrix multiplication.
- Adds `packed_modules_model_mapping` for Kimi-K2-Thinking model and
processing logic for `weight_packed` field.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: zhoux77899 <zhouxiang100@huawei.com>
Signed-off-by: Ruri <33858552+zhoux77899@users.noreply.github.com>
Signed-off-by: Ruri <zhouxiang100@huawei.com>
### What this PR does / why we need it?
Support pooling models (like `bge-reranker-v2-m3`) in vllm-ascend, this
pr covered the three model types of embed (cls_token, mean_token,
lasttoken).
After this
[commit](17373dcd93),
vllm has provided support for adapting pooling models on the v1 engine.
This PR includes corresponding adaptations on the vllm-ascend side.
Fixes#1960
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: lianyibo <lianyibo1@kunlunit.com>
Signed-off-by: MengqingCao <cmq0113@163.com>
Co-authored-by: MengqingCao <cmq0113@163.com>
aclgraph is stable and fast now. Let's drop torchair graph mode now.
TODO: some logic to adapt torchair should be cleaned up as well. We'll
do it in the following PR.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
### What this PR does / why we need it?
As support for the mooncake connector is now available, the llmdatadist
connector is no longer being maintained, so the llmdatadist-related
files need to be retired.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
By ci
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: wangxiaoteng <wangxiaoteng@huawei.com>
Signed-off-by: liziyu <liziyu16@huawei.com>
Co-authored-by: liziyu <liziyu16@huawei.com>
### What this PR does / why we need it?
This PR is to fix a smoking test failure. Adjust mtp_proposer and
model_runner_v1 to route MTP decoding through the non‑fused MoE
implementation while keeping the overall inference flow unchanged.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: mojave2 <chenchen145@huawei.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
### What this PR does / why we need it?
Support pp for kv pool
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: baxingpiaochong <771405853@qq.com>
### What this PR does / why we need it?
In reinforcement learning scenarios, the current inference applies a
transpose operation to the weights. For a cleaner architecture, the
weight transpose module was moved to wakeup.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: lhp-deep <liuhaopeng1@huawei.com>
Co-authored-by: weijinqian0 <1184188277@qq.com>
### What this PR does / why we need it?
Fix dp padding logic in dummyrun. After
https://github.com/vllm-project/vllm/pull/28579, `num_tokens` will be
padded in `CudagraphDispatcher`, thus we also need to do the pad in the
dummy_run.
### How was this patch tested?
Test locally with the following scripts
```bash
VLLM_USE_MODELSCOPE=true python3 -m vllm.entrypoints.openai.api_server \
--model wemaster/deepseek_mtp_main_random_bf16 \
--trust-remote-code \
--data-parallel-size 4 \
--tensor-parallel-size 1 \
--compilation-config '{"cudagraph_capture_sizes":[96],"cudagraph_mode":"FULL_DECODE_ONLY"}' \
--enable-expert-parallel
```
```bash
vllm bench serve --model wemaster/deepseek_mtp_main_random_bf16 --endpoint /v1/completions --dataset-name random --random-input 512 --random-output 100 --num-prompts 48 --request-rate 1 --ready-check-timeout-sec 0
```
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
Move the logic for adjusting ACL graph capture sizes for speculative
decoding from the generic utility module into a dedicated method within
the compilation configuration.
This change improves code organization and encapsulation by making the
compilation configuration responsible for managing its own state. The
model runner now triggers this adjustment directly, providing the
necessary context.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
None.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
this pr aims to support async_scheduling for mtp, which refer to vllm pr
https://github.com/vllm-project/vllm/pull/24799.
and this pr fix some synchronize problem in vllm-ascend.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>
It's safe to drop ascend scheduler now. The related test and doc has
been removed already
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
This PR introduces the Ascend implementation of the
`dispatch_ffn_combine` kernel and wires it into the vLLM-Ascend runtime,
together with follow‑up fixes to ensure the kernel builds and runs
correctly in CI.
- Add full host and device implementation of the `dispatch_ffn_combine`
kernel under `csrc/dispatch_ffn_combine`, including tiling logic, MOE
routing helpers, and kernel utilities for quantized FFN dispatch.
- Integrate the new kernel with the PyTorch binding
(csrc/torch_binding.cpp, csrc/torch_binding_meta.cpp) and the Ascend
runtime (vllm_ascend/ascend_forward_context.py,
vllm_ascend/worker/model_runner_v1.py).
- Extend fused MoE communication and token dispatch support in
`vllm_ascend/ops/fused_moe`, adding methods/utilities needed by the new
dispatch path.
- Update quantization logic in vllm_ascend/quantization/w8a8_dynamic.py
to support the new FFN dispatch flow.
- Fix kernel build issues by adjusting `csrc/build_aclnn.sh`, CMake
configuration, and include/namespace usage in the new kernel files.
- Add an end‑to‑end nightly test
`tests/e2e/nightly/ops/test_dispatch_ffn_combine.py` and helper
utilities in `vllm_ascend/utils.py` to validate the new kernel.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.12.0
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.12.0
---------
Signed-off-by: mojave2 <chenchen145@huawei.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
Due to the requirement of the FIA operator that the **query.shape[0]**
must match **actual_seq_len[-1]**, in graph mode and multi-DP scenarios,
the query is padded to the size of **num_input_token**. This leads to
validation errors during tiling in the operator. However, since the
padding is applied at the end of the query, it does not affect the
actual execution result of the operator, and the precision remains
unaffected.
<img width="2434" height="49" alt="image"
src="https://github.com/user-attachments/assets/63520816-fbc3-4382-82b9-89dbb1492f6c"
/>
Our modification padding both **actual_seq_len** and
**actual_seq_len_kv** to resolve the validation issue in the operator.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.11.2
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2
Signed-off-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
Co-authored-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
### What this PR does / why we need it?
Provide high-performance AscendC operators lightning_indexer and
sparse_flash_attention to boost the execution performance of the
DeepSeek v3.2 model. Meanwhile, adapt the two AscendC operators to
vllm-ascend framework.
### Does this PR introduce _any_ user-facing change?
No (only underlying operator optimizations, with no user-facing changes)
### How was this patch tested?
- vLLM version: v0.11.2
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2
Signed-off-by: MingYang119 <songmingyang@huawei.com>
### What this PR does / why we need it?
This PR adds a triton rope kernel witch supports scenarios of `rope_dim
!= head_dim`. This can save the split op before rope and the concat op
after rope. Profiling shows improvement.
### Does this PR introduce _any_ user-facing change?
None
### How was this patch tested?
I will add related ut after ci integrated with triton.
- vLLM version: v0.11.2
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2
---------
Signed-off-by: whx-sjtu <2952154980@qq.com>
### What this PR does / why we need it?
This PR integrate suffix decoding (https://arxiv.org/abs/2411.04975)
from vllm (https://github.com/vllm-project/vllm/pull/25784)
#
Suffix Decoding is a dynamic n-gram matching method that:
1. Uses suffix trees to generate speculative tokens quickly using branch
frequency counts.
2. Can keep a history of prior model responses, which tends to work very
well with repetitive agentic use cases.
3. Can be dynamically updated with newly generated tokens, and FIFO
eviction of older requests.
#
### Does this PR introduce _any_ user-facing change?
This feature should be implemented as opt-in and remain seamless for
users who do not require suffix speculative decoding.
For users who wish to enable it, they must first install
arctic-inference:
`pip install arctic-inference
`
After installation, the suffix speculative decoding feature can be
enabled using the following speculative config:
`--speculative_config '{"method": "suffix", "num_speculative_tokens":
5}'
`
### How was this patch tested?
This PR is currently being tested on vLLM
main:83f478bb19
with PR https://github.com/vllm-project/vllm/pull/25784
In our previous testing, suffix decoding achieved a 13%-30% throughput
improvement over n-gram on the sonnet dataset, tested on vllm-ascend
v0.9.1 with concurrency ranging from 2 to 40.
- vLLM version: v0.11.2
---------
Signed-off-by: fluctlux <38945811+fluctlux@users.noreply.github.com>
### What this PR does / why we need it?
Previously, the dummy run executed compute_logits only once, regardless
of num_speculative_tokens. This caused execute_model to hang on
compute_logits when lm head tensor parallelism exceeded 1. The fix
ensures compute_logits executes correctly during dummy run, matching
num_speculative_tokens.
I set the `non_blocking` argument to False when moving
`exceeds_max_model_len` to the CPU. From what I understand, using
`non_blocking=True` and immediately accessing the tensor on the CPU can
cause accuracy problems. However, this issue doesn't happen when
transferring data to a device. ref:
https://discuss.pytorch.org/t/should-we-set-non-blocking-to-true/38234/18
- vLLM version: v0.11.0
- vLLM main:
2918c1b49c
---------
Signed-off-by: Jade Zheng <zheng.shoujian@outlook.com>
Ascend scheduler was added for non chunk prefill case before, since that
the npu ops didn't work well with chunked prefill.
Now the ops with chunked prefill work better, it's time to remove the
ascend scheduler to use vLLM default scheduler.
- vLLM version: v0.11.2
---------
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Fix model run _npu_flash_attention in _forward_prefill_no_cache hang
issue, it was caused by wrong attention mask dtype.
### How was this patch tested?
Yes, tesed on Qwen2.5-VL and Qwen2.5-Omni
- vLLM version: v0.11.0
- vLLM main:
2918c1b49c
Signed-off-by: Ting FU <futing10@huawei.com>
### What this PR does / why we need it?
Currently, there are two paths to judge the chip type in code,
`get_ascend_soc_version` use `get_soc_version` api in torch_npu, and
`is_310p` `use _build_info.__soc_version__`, which generate when
install. We need to unify the two paths.
We need to unify these codes based on the following points:
1. We need to ensure consistency in chip type judgment between compiling
and running states;
2. In compiling state, we need chip type to complete op's compilation,
but in running state, we only need device
type(910B/910_93/310P/910_95/etc) to make code branch judgement;
3. In compiling state, torch_npu may not have been installed yet, so we
can't use torch_npu's api.
Based on the above points, we have made the following changes:
1. When user set env `SOC_VERSION`, use it; when not set, query
soc_version by `npu-smi`;
2. generate device_type based on soc_version when compiling, and write
`__device_type__` instead of `__soc_version__` in `_build_info.py`;
3. In running state, use `__device_type__` to judge code branch.
### Does this PR introduce _any_ user-facing change?
When not set env `SOC_VERSION`, it will not be `ASCEND910B1` by default,
we will query soc_version by `npu-smi`. And env `SOC_VERSION` must be in
the list `soc_to_device` in `setup.py`.
- vLLM version: v0.11.0
- vLLM main:
2918c1b49c
Signed-off-by: zzzzwwjj <1183291235@qq.com>
### What this PR does / why we need it?
When cudagraph_mode is set to FULL_DECODE_ONLY, if dp > 1, the dummy-run
process will be triggered. When calling the update_attn_params function,
the num_tokens parameter needs to be passed, and this value is obtained
through positions.shape[0]. However, the multimodal model uses mRope
(multi-dimensional rotary positional embeddings), which causes the shape
of positions to be 2. As a result, the value obtained from
positions.shape[0] is incorrect. We solve this problem by replacing
positions.shape[0] with num_tokens.
### Does this PR introduce _any_ user-facing change?
NO
### How was this patch tested?
vLLM version: v0.11.0rc3
vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
- vLLM version: v0.11.0
- vLLM main:
2918c1b49c
---------
Signed-off-by: wujinyuan1 <wjy9595@qq.com>
Co-authored-by: wujinyuan1 <wjy9595@qq.com>
### What this PR does / why we need it?
vllm-ascend need to dump data during model execution to debug some
precision problems, here msprobe provide the corresponding abilities, so
msprobe will join vllm-ascend to make debug easier
### Does this PR introduce _any_ user-facing change?
```
'dump_config': '/path/to/config.json'
```
- vLLM version: v0.11.0
- vLLM main:
2918c1b49c
---------
Signed-off-by: Tjh-UKN <2559659915@qq.com>
The main purposes of this PR are as follows:
1. Remove the multicast-related code;
Reason:
1. In the scenario like a2 Dual-System Back-to-Back Networking，the
performance is worse than all_gather. Before the modification, in e2e
test, it was 3 tps; after the modification, it is 10 tps.
2. At the same time, we usually enable the SP feature，it is consistent
with the current logic.
3. The advantage of broadcast communication lies in the fact that it
does not suffer from uneven DP load and does not require the prefill ACL
graph to be enabled. But we support prefill Acl graph recently.
So we think there is no need to maintain the multicast as one choice in
moe communication.
Performance benefits are as follows:
When not enable_flashcomm1, TTFT remains relatively stable at around
43000ms, which is approximately 15000ms faster than before the
modification.
When enable_flashcomm1, there is no diffenence, TTFT remains relatively
stable at around 29000ms.
- vLLM version: v0.11.0
- vLLM main:
2918c1b49c
---------
Signed-off-by: weijinqian_v1 <weijinqian@huawei.com>
Signed-off-by: weijinqian0 <1184188277@qq.com>
Co-authored-by: weijinqian_v1 <weijinqian@huawei.com>
There is a lot hack code for v0.11.0, which makes the code hard to
upgrade to newer vLLM version. Since v0.11.0 will release soon. Let's
drop v0.11.0 support first. Then we'll upgrade to v0.11.2 soon.
- vLLM version: v0.11.0
- vLLM main:
2918c1b49c
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
In [#26016](https://github.com/vllm-project/vllm/pull/26016), vllm
change the `cudagraph_capture_sizes` to be in ascending order. This PR
fixes related issues caused by this.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
- vLLM version: v0.11.0
- vLLM main:
2918c1b49c
---------
Signed-off-by: Angazenn <supperccell@163.com>
### What this PR does / why we need it?
Support the Qwen3-Next-80B-A3B-Instruct quantization model and Fix the
NZ issue. Triton kernel doesn't support data format nz, thus we skip
converting weight to nz on layer `conv1d`
- vLLM version: v0.11.0
- vLLM main:
2918c1b49c
---------
Signed-off-by: IncSec <1790766300@qq.com>
### What this PR does / why we need it?
Currently, the MTP model still runs in eager in full graph mode. This PR
adapts the MTP with the full graph capture and execution. When the graph
mode is set to "FULL_DECODE_ONLY", the MTP will run in full-graph to
improve the performance.
The change in both disable_padded_drafter_batch is True and False case
include:
1. Add _mtp_graph_params in acl_graph.py to isolate the data of main
model and the data of MTP.
2. Padding some metadata in mla_v1.py when in fullgraph mode.
3. Fixed the essential data address that will be used in model.forward.
4. Adapted according to the aclgraph capture framwork:
1). Rebuild MTP model with ACLGraphWrapper.
2). Add common attn metadata when start capture in MTP dummy_run.
3). Add common attn metadata update in MTP.
4). Addapted data update when num_speculative_tokens > 1.
5. Add a patch of MTP to adapt vllm v0.11.0.
Existing Issues:
1. When disable_padded_drafter_batch=True and running in FullGraph mode,
the data of the first-round requests in MTP is abnormal. We need to
identify the cause subsequently.
2. When disable_padded_drafter_batch=False and running in FullGraph
mode, the acceptance rate of the second and third tokens will decrease
(For example, if we set the num_speculative_tokens=3, the acceptance
rate of first token is 90%, the second is only 50% lower than 60%, the
third is only 20% lower than 30%). The reason is that the data processed
after the model runs does not match. This is a problem from another PR.
It works fine in eager and PIECEWISE mode, but has problem in FullGraph
mode. Once we have a solution, we will submit a bugfix.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.11.0
- vLLM main:
2918c1b49c
---------
Signed-off-by: anon189Ty <Stari_Falcon@outlook.com>
