…(#7603)
### What this PR does / why we need it?
Block verify uses cumprod(target_probs / draft_probs) for joint
acceptance. Suffix/ngram methods have
draft_probs=None, the fallback draft_token_probs=1.0 with cumprod is not
equivalent to per-token
verification, causing incorrect accept/reject results. Fix:
using_block_verify = max_spec_len >= 3 and draft_probs is not None.
MTP/Eagle3 unaffected.
- vLLM version: v0.18.0
- vLLM main:
ed359c497a
<!-- Thanks for sending a pull request!
BEFORE SUBMITTING, PLEASE READ
https://docs.vllm.ai/en/latest/contributing/overview.html
-->
### What this PR does / why we need it?
<!--
- Please clarify what changes you are proposing. The purpose of this
section is to outline the changes and how this PR fixes the issue.
If possible, please consider writing useful notes for better and faster
reviews in your PR.
- Please clarify why the changes are needed. For instance, the use case
and bug description.
- Fixes #
-->
### Does this PR introduce _any_ user-facing change?
<!--
Note that it means *any* user-facing change including all aspects such
as API, interface or other behavior changes.
Documentation-only updates are not considered user-facing changes.
-->
### How was this patch tested?
<!--
CI passed with new added/existing test.
If it was tested in a way different from regular unit tests, please
clarify how you tested step by step, ideally copy and paste-able, so
that other reviewers can test and check, and descendants can verify in
the future.
If tests were not added, please describe why they were not added and/or
why it was difficult to add.
-->
Signed-off-by: liuchenbing <chenliumail@163.com>
Co-authored-by: liuchenbing <chenliumail@163.com>
### What this PR does / why we need it?
In order to adapt to the GLM model, logits were passed in the sample,
which can cause accuracy issues in version 0.15.0.
- vLLM version: v0.15.0
- vLLM main:
83b47f67b1
Signed-off-by: GDzhu01 <809721801@qq.com>
### What this PR does / why we need it?
This PR aims to update `target_probs` to `target_logits` in
`rejection_sample`, for catching up with
https://github.com/vllm-project/vllm/pull/32852. Otherwise, sampling
with temperature will incur accuracy problem where tokens can be
accepted or rejected unreasonably.
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
by ci
- vLLM version: v0.15.0
- vLLM main:
13397841ab
Signed-off-by: Zetong Li <slippersss@126.com>
### What this PR does / why we need it?
1. MagicMTP (paper: "Block Verification Accelerates Speculative
Decoding") was introduced to consider the influence among multiple draft
tokens, improving the acceptance rate without compromising accuracy.
2. Added Triton and PyTorch implementations, and added E2E test cases.
### Does this PR introduce _any_ user-facing change?
MagicMTP will automatically take effect when the parameter
"num_speculative_tokens" >= 3.
- vLLM version: v0.13.0
- vLLM main:
7157596103
Signed-off-by: chenaoxuan <cax1165@163.com>
### What this PR does / why we need it?
Import global var form vllm instead of overwirte it, so that we could
use the correct global variant value
- vLLM version: v0.13.0
- vLLM main:
5326c89803
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
This PR moves reject sample related triton kernels into `ops/triton`.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
CI passed with existing test.
- vLLM version: release/v0.13.0
- vLLM main:
5fbfa8d9ef
---------
Signed-off-by: whx-sjtu <2952154980@qq.com>
### What this PR does / why we need it?
#4443 introduces a precision issue in scenarios where MTP >= 3 + deepseek v3.1, and this pr reverts it
- vLLM version: release/v0.13.0
- vLLM main:
bc0a5a0c08
Signed-off-by: GDzhu01 <809721801@qq.com>
### What this PR does / why we need it?
1. MagicMTP (paper: "Block Verification Accelerates Speculative
Decoding") was introduced to consider the influence among multiple draft
tokens, improving the acceptance rate without compromising accuracy.
2. The rejection sampling logic in rejection_sampler.py was restructured
using Triton-Ascend, enabling it to operate under high concurrency, thus
resolving CPU and NPU operator bottlenecks and enhancing throughput.
### Does this PR introduce _any_ user-facing change?
MagicMTP will automatically take effect when the parameter
"num_speculative_tokens" >= 3.
- vLLM version: v0.11.2
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2
Signed-off-by: chenaoxuan <cax1165@163.com>
### What this PR does / why we need it?
1. Use optimized apply_top_k_top_p for NPU platfrom in rejection
sampler; (avoid scatter elements which can reduce ~26ms TPOT with bs=24
per DP)
2. <del>Avoid D2H Synchronization before calling npu_top_k_top_p
introduced by parameter validation which improves inference speed with
`async_scheduling` enabled;</del> In order to elminate the D2H
synchronization introduced by parameter validation before calling
`npu_top_k_top_p`, we directly drop this fused operator since the
performance improvement is not significant compared to async_scheduling
and may bring potential accuracy problem.
3. Refactor the implementation of AscendTopKTopPSampler to align that of
vLLM.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
E2E serving test with combinations of `k=500` and `p=0.95` with
async_scheduling in single node and wide-EP scenarios.
- vLLM version: v0.11.0
- vLLM main:
83f478bb19
---------
Signed-off-by: linfeng-yuan <1102311262@qq.com>
Co-authored-by: realliujiaxu <realliujiaxu@163.com>
### What this PR does / why we need it?
Add top_p,top_k in EAGLE e2e
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: zhaomingyu <zhaomingyu13@h-partners.com>
### What this PR does / why we need it?
This PR introduces optimized Triton implementations for the
rejection_greedy_sample_kernel and expand_kernel, delivering superior
performance compared to the existing Triton implementations. The new
Triton kernels maintain full functional accuracy while delivering
significant performance improvements across various batch sizes and MTP
configurations.
### Does this PR introduce _any_ user-facing change?
Yes, this PR modifies rejection_sampler.py to use optimized Triton
kernels:
- rejection_greedy_sample_kernel is enhanced with
rejection_greedy_sample_spec_len_1_triton and
rejection_greedy_sample_triton implementations
- expand_kernel receives a performance-optimized Triton version
These changes provide substantial performance improvements while
maintaining backward compatibility
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: yuxingcyx <yuxingchen.math@gmail.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
Currently, we are using `AscendRejctionSampler` that extends from
`RejctionSampler` in spec decoding. `AscendRejctionSampler` override
`forward` of `RejctionSampler`, only aming to replace `rejection_sample`
func. This
causes a lot of code of `RejctionSampler` cannot be reused, for example:
- https://github.com/vllm-project/vllm/pull/19482
- https://github.com/vllm-project/vllm/pull/26060
- https://github.com/vllm-project/vllm/pull/29223
#### Proposed Change:
- Delete `AscendRejctionSampler` and use `RejctionSampler` directly in
model runner.
- Patch `RejctionSampler.expand_batch_to_tokens` and
`RejctionSampler.rejection_sample`, maybe a better way is to make them
as custom ops.
- Modify `NPUModelRunner` following
https://github.com/vllm-project/vllm/pull/26060
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- [x] test logits processor for spec decoding
- [x] test logprobs for spec decoding
- [x] test logprobs for spec decoding + async shcheduling (test with
https://github.com/vllm-project/vllm-ascend/pull/4893/)
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: realliujiaxu <realliujiaxu@163.com>
…
### What this PR does / why we need it?
When speculative decoding is enabled and temperature > 0, bonus_logits
and target_logits are sampled separately:
1. bonus_logits are sampled using a fused torch_npu.npu_top_k_top_p
operator invoked inside the main sampler,
2. while target_logits are sampled within the rejection sampler using a
less-optimized implementation composed of smaller operators.
Consequently, the cumsum operation in the top-p sampling for
target_logits becomes especially time-consuming, leading to performance
degradation.
<img width="1029" height="623" alt="image"
src="https://github.com/user-attachments/assets/1969f561-6aa5-41b3-9a87-1f64d4321cbf"
/>
Apply the fused operator to the sampling of target_logits as well to
reduce overhead
<img width="1039" height="572" alt="image"
src="https://github.com/user-attachments/assets/1e6563da-3418-405d-b657-7bbe10dd0924"
/>
### 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: funanyang <985619145@qq.com>
Co-authored-by: weijinqian0 <1184188277@qq.com>
### What this PR does / why we need it?
Original `sample_recover_tokens_kernel` of reject sampler didn't tile
the vocab size dim, whitch will cause ub overflow problem for models
with big vocab size like deepseek. This PR adds tiling to the vocab size
dim to avoid this problem.
Note that currently we just use a emperical `SUB_BLOCK_SIZE` of `4*1024`
for functionality. If in the future this kernel becomes performance
bottle neck, we can use triton autotune to optimize this. What's more,
we have to disable multibuffer of this kernel due to some accuracy
issues.
- vLLM version: v0.12.0
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.12.0
Signed-off-by: whx-sjtu <2952154980@qq.com>
Co-authored-by: weijinqian0 <1184188277@qq.com>
### What this PR does / why we need it?
Replace pyorch implement of sampling with triton kernels
### Does this PR introduce _any_ user-facing change?
No
- vLLM version: v0.11.2
---------
Signed-off-by: Lord_of_Ironhill <suiweiyi@huawei.com>
Signed-off-by: whx-sjtu <2952154980@qq.com>
Co-authored-by: Lord_of_Ironhill <suiweiyi@huawei.com>
Co-authored-by: whx-sjtu <2952154980@qq.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?
Replace masked in-place assignment with a device-side torch.where so
selection stays on-device, allowing subsequent device ops to be enqueued
earlier and removing an implicit D2H sync, reducing latency by several
hundreds μs on Ascend.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
None.
- vLLM version: v0.11.0
- vLLM main:
83f478bb19
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
This PR port optimization in PR #2002 to main and makes it cleaner.
- vLLM version: v0.10.0
- vLLM main:
afa5b7ca0b
---------
Signed-off-by: whx-sjtu <2952154980@qq.com>
### What this PR does / why we need it?
Support MTP with：
- [x] V0 Scheduler
- [x] TorchAir
- [x] Single DP
- [x] Multi DP
- [x] Disaggregate PD
Known issues：
- [ ] Not support V1 Scheduler (chunked prefill), will be supported in a
few weeks
- [ ] vllm v0.10.0 does not support metrics with `DP > 1` right now,
need to comment out the line 171-175 in file
`vllm/vllm/v1/metrics/loggers.py`
```
if (len(self.engine_indexes) > 1
and vllm_config.speculative_config is not None):
raise NotImplementedError("Prometheus metrics with Spec Decoding "
"with >1 EngineCore per AsyncLLM is not "
"supported yet.")
```
To start an online server with torchair enabled, here is an example:
```
python -m vllm.entrypoints.openai.api_server \
--model="/weights/DeepSeek-R1_w8a8/" \
--trust-remote-code \
--max-model-len 40000 \
--tensor-parallel-size 4 \
--data_parallel_size 4 \
--max-num-seqs 16 \
--no-enable-prefix-caching \
--enable_expert_parallel \
--served-model-name deepseekr1 \
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}' \
--quantization ascend \
--host 0.0.0.0 \
--port 1234 \
--additional-config '{"ascend_scheduler_config":{"enabled":true,"enable_chunked_prefill":false},"torchair_graph_config":{"enabled":true,"graph_batch_sizes":[16]},"enable_weight_nz_layout":true}' \
--gpu_memory_utilization 0.9
```
offline example with torchair enabled
```
from vllm import LLM, SamplingParams
prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
]
# Create a sampling params object.
sampling_params = SamplingParams(max_tokens=16, temperature=0)
# Create an LLM.
llm = LLM(
model="/home/data/DeepSeek-R1_w8a8/",
tensor_parallel_size=16,
max_num_seqs=16,
gpu_memory_utilization=0.9,
distributed_executor_backend="mp",
enable_expert_parallel=True,
speculative_config={
"method": "deepseek_mtp",
"num_speculative_tokens": 1,
},
trust_remote_code=True,
enforce_eager=False,
max_model_len=2000,
additional_config = {
'torchair_graph_config': {
'enabled': True,
"graph_batch_sizes": [16],
'enable_multistream_shared_expert': False,
},
"ascend_scheduler_config": {
"enabled": True
},
# 'expert_tensor_parallel_size': 16,
}
)
# Generate texts from the prompts.
# llm.start_profile()
outputs = llm.generate(prompts, sampling_params)
# llm.stop_profile()
for output in outputs:
prompt = output.prompt
generated_text = output.outputs[0].text
print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
```
- vLLM version: v0.10.0
- vLLM main:
302962e806
---------
Signed-off-by: xuyexiong <xuyexiong@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?
<!--
- 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 #
-->
Make spec decode support for V1 Engine
- Currently, Ascend does not support the triton kernel. PyTorch is used
to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is
not as good as Triton. Therefore, ascend c is used to implement the
function in the future.
- Currently, spec decode supports only the ngram algorithm. The eagle
algorithm needs to be further adapted.
### 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.
-->
Not change user facing.
### 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.
-->
test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and
`tests/sample/test_rejection_sampler.py`, test base function of
rejection sampler and e2e function of spec decode.
Signed-off-by: ponix-j <657511300@qq.com>
