### What this PR does / why we need it?
Some bug fixes, mainly including:
1. For A2, the number of experts each single card cannot be greater than
16 when using MC2. The PR fixed the error in the A2 moe communication
method selection, which would cause the selection of an incorrect
communication method when the number of model experts exceeds 256. For
example, when using an A2 16-cards model to load the PD-disaggregation D
node with Qwen3.5 series models, the incorrect MC2 method would be
chosen.
2. Fixed the issue where the layerwise connector sends the kv-cache of
the MTP layer multiple times when `num_spec_tokens` > 1. Now, the
kv-cache is sent only when the MTP layer is forward for the first time.
3. Fix the accuracy issue of qwen3.5 when using MTP for PD
disaggregation. The cause is that `num_decode_draft_tokens` does not
consider that `spec_tokens` are not existed during the first inference
when PD disaggregation (`spec_tokens` are generated during the first
inference). However, `spec_tokens_padding` is added by
`recomputed_scheduler`. As a result, `gdn_metadata` incorrectly
considers that the prefill with a length of 2 is performed.
---------
Signed-off-by: nwpu-zxr <zhouxuerong2@huawei.com>
Signed-off-by: zxr2333 <64738772+nwpu-zxr@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.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?
Two problems have been solved in this pr.
These problems occur in the `FULL_DECODE_ONLY` mode that `num_tokens`
should be padded to some value in `cudagraph_capture_sizes`.
1. We found the length of `seq_lens_list` in drafter's `attn_metadata`
is 1 shorter than expected. It will raise a kernel exception to make
vllm crash.
e.g., `num_reqs` = 3, `cudagraph_capture_sizes` = [20],
`actual_seq_lengths_q` is padded well to [4, 8, 12, 20]. But
`seq_lens_list` = [5742, 4700, 7996], it is not padded.
3. Though the length of `seq_lens_list` in target's `attn_metadata` is
the same as expected in `FULL_DECODE_ONLY`, some data are corrupted at
the end of the list.
e.g., `num_reqs` = 3, `cudagraph_capture_sizes` = [20],
`actual_seq_lengths_q` is padded well to [4, 8, 12, 20]. But
`seq_lens_list` = [5742, 4700, 7996, 5738], it has corrupted at the end
of the list.
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
Signed-off-by: drslark <slarksblood@qq.com>
### What this PR does / why we need it?
This PR restores #7029, which adds W8A8C8 support for dsv3.2/glm5 using
the `lightning_indexer_quant` ops in the pd-mix stage.
The original PR was reverted by #7288 because the patch did not work
with the recompute scheduler.
This PR also fixes the patching issue so that it works correctly with
the recompute scheduler.
### Does this PR introduce _any_ user-facing change?
Yes. To enable LI C8, users need to set the `enable_sparse_c8` option to
`"true"` in `additional_config`.
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: rjg-lyh <1318825571@qq.com>
### What this PR does / why we need it?
This reverts commit 7ed9e9de69, which
introduces an issue that the patch doesn't work with recompute scheduler
enabled.
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
To support prefix cache for Qwen3.5/Next in vLLM-Ascend, this PR mainly
follows the design in
[#30877](https://github.com/vllm-project/vllm/pull/30877) and inherits
changes to functions which are overridden in vLLM-Ascend.
Note:
1. `--mamba-cache-mode align` && PD disaggregation is still not
supported yet in vLLM v0.17.0(see
https://github.com/vllm-project/vllm/blob/main/vllm/v1/core/sched/scheduler.py#L295).
2. The current implementation of hybrid kv cache might result in a very
large block_size when scheduling. For example, if we run Qwen3.5-35B-A3B
with `-tp 2`, the block_size is adjusted to 2048, which means that any
prefix shorter than 2048 will never be cached. Although this behavior is
consistent with vLLM, it still needs improvements in the future.
3. `--mamba-cache-mode align` requires to copy mamba states during
forward steps. vLLM uses a triton kernel to implement it. However, the
original version run into some bugs on Ascend hardwares. Thus we patch a
new triton kernel to avoid this bug.
### Does this PR introduce _any_ user-facing change?
To use mamba prefix cache, set `--enable-prefix-caching` and
`--mamba-cache-mode align`. Note that the mamba state copy function(see
[do_mamba_copy_block](https://github.com/vllm-project/vllm/blob/main/vllm/v1/worker/mamba_utils.py#L132))
does not provide a torch native version, thus it might have trouble if
users can't use triton.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: Angazenn <supperccell@163.com>
### What this PR does / why we need it?
This pr fix the Unified draft parallel feature.
1. In Draft model proposer, there are exceed 1 attention layers in
target model, thus removing the assertion on layer number.
2. we should get block size through `draft_attn_groups` instead of
`attn_metadata_builder` after 0.17.0.
3. `attn_update_stack_num_spec_norm` shouldn't be done when unified
draft parallel is enabled
### How was this patch tested?
Test pass with
`tests/e2e/singlecard/spec_decode/test_v1_spec_decode.py::test_parallel_drafting_acceptance`,
which is already included in CI
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
Drop 0.16.0 support in main
- Fix eagle proposer break introduced by
https://github.com/vllm-project/vllm/pull/34552. Mainly change to use
the draft attention group to initialize the attention metadata builder.
- Fix the `ModelRunner` has no attribute `cudagraph_capture_sizes`
error, which is a bug in vLLM v0.17.0, and fixed by a later pr
https://github.com/vllm-project/vllm/pull/30515
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
This PR supports W8A8C8 in dsv3.2/glm5 with lightning_indexer_quant ops
in pd-mix stage mainly.
Because the code for the current PD-disaggregated scenario is still
under refactoring and cleanup, this PR prioritizes ensuring the C8
functionality in the pd-mix scenario.
The next steps are planned in two parts:
① Once the optimized scatter operator is updated, we will replace the
original operator to improve the performance of storing k_scale.
② Once the code logic for the PD-disaggregated scenario becomes stable,
we will carry out more comprehensive validation and make appropriate
adaptations.
③ Because enabling C8 currently introduces several new operators whose
performance still needs improvement, performance may regress in some
scenarios. Therefore, only after all the operators are fully ready can
we ensure that this feature does not cause any performance degradation.
At that point, we will enable this feature by default and remove the
switch in `additional_config`.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
CI passed with new added/existing test.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: rjg-lyh <1318825571@qq.com>
### What this PR does / why we need it?
This PR fixes a bug in `reshape_kvcache_tensors` when reshaping the
Mamba cache for models like Qwen3.5. The previous implementation did not
correctly handle cases where the KV cache tensors have different data
types. This change ensures that slicing is performed based on byte
offsets before reshaping the tensors, which correctly handles
heterogeneous dtypes.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
By CI.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
Signed-off-by: nwpu-zxr <zhouxuerong2@huawei.com>
### What this PR does / why we need it?
The merged graph of draft in `FULL` mode is broken now.
This pr solves it.
Also, `actual_seq_lengths_q` in `model_runner` is found redundant, so,
it is removed.
It depends on https://github.com/vllm-project/vllm-ascend/pull/7144 and
https://github.com/vllm-project/vllm-ascend/pull/7148.
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
Test code is shown as below:
```python
prompts = [
"1.Who are you?",
"2. Who are you?",
]
sampling_params = SamplingParams(temperature=0.0, top_p=0.95, top_k=40, max_tokens=200)
llm = LLM(
model="/home/some-model/Meta-Llama-3.1-8B-Instruct",
tensor_parallel_size=1,
max_num_seqs=32,
# enforce_eager=True,
disable_log_stats=False,
distributed_executor_backend="mp",
gpu_memory_utilization=0.7,
async_scheduling=True,
speculative_config={
"enforce_eager": True,
"model": "/home/some-model/EAGLE3-LLaMA3.1-Instruct-8B",
"disable_padded_drafter_batch": False,
"method": "eagle3",
"num_speculative_tokens": 3,
},
compilation_config={
"cudagraph_mode": "FULL",
"cudagraph_num_of_warmups": 1,
},
max_model_len=4096,
enable_prefix_caching=False,
)
outputs = llm.generate(prompts, sampling_params)
```
The result before:
```text
File "/vllm-workspace/vllm-ascend/vllm_ascend/attention/attention_v1.py", line 575, in full_graph_fia
graph_params.events[num_tokens].append(event)
~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^
KeyError: 132
```
The result after:
```text
--------------------------------------------------
total_num_output_tokens: 400
num_drafts: 242
num_draft_tokens: 726
num_accepted_tokens: 156
mean acceptance length: 1.64
--------------------------------------------------
acceptance at token 0: 0.42
acceptance at token 1: 0.16
acceptance at token 2: 0.07
```
We also test `FULL_DECODE_ONLY` mode.
The result is:
```text
--------------------------------------------------
total_num_output_tokens: 400
num_drafts: 244
num_draft_tokens: 732
num_accepted_tokens: 155
mean acceptance length: 1.64
--------------------------------------------------
acceptance at token 0: 0.42
acceptance at token 1: 0.16
acceptance at token 2: 0.06
```
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
Signed-off-by: drslark <slarksblood@qq.com>
### What this PR does / why we need it?
Fix the issue where no exception is thrown when graph capture fails.
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
Signed-off-by: WithHades <244036962@qq.com>
### What this PR does / why we need it?
related to vllm PR #34043 this pr delete func
‘relax_for_mixed_batch_cudagraphs’, num_reqs no longer equals the actual
number of requests, due to fia operator requires that
query_start_loc[-1] equals the total number of computed tokens, so this
func delete cause the ifa error.
In full graph mode, set num_reqs_paded = num_reqs to fix the error
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
Co-authored-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
### What this PR does / why we need it?
Fix acceptance and high-concurrency bug in eagle3 and cp enabled
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
tests and ut
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: lilinsiman <lilinsiman@gmail.com>
### What this PR does / why we need it?
Fix the LoRA e2e test accuracy issue that introduced by the upstream PR
https://github.com/vllm-project/vllm/pull/32005
### How was this patch tested?
pytest -sv tests/e2e/singlecard/test_llama32_lora.py
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
---------
Signed-off-by: paulyu12 <507435917@qq.com>
Signed-off-by: yupeng <507435917@qq.com>
## Summary
- Move `_update_states_after_model_execute` call from after main model
sampling to after draft model execution
- This reordering reduces pipeline bubbles between main model and draft
model execution
- No accuracy impact - the state update operation is independent of
draft token proposal
## Performance Impact
Reduces idle time between main model and draft model execution stages,
improving overall MTP (Multi-Token Prediction) performance.
- vLLM version: v0.15.0
- vLLM main:
83b47f67b1
---------
Signed-off-by: SlightwindSec <slightwindsec@gmail.com>
Co-authored-by: wanghuanjun2113 <wanghuanjun2113@gmail.com>
### What this PR does / why we need it?
Supports contiguous tensor hybrid-attn kv-cache on fullattn-mamba hybrid
model, such as Qwen3Next and Qwen3.5.
Due to the restrictions of Ascend operators, all KV tensors, conv
tensors, and SSM tensors must be contiguous. Therefore, this PR uses the
following solution to generate the KV cache:
tensor1: [(kv_padding), conv , ...]
tensor2: [k , ssm , ...]
tensor3: [v , (mamba_padding), ...]
Under this scheme, although some waste may occur, the tensors of all
caches are guaranteed to be contiguous.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
By CI.
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
---------
Signed-off-by: nwpu-zxr <zhouxuerong2@huawei.com>
### What this PR does / why we need it?
The index-select operation `mrope_positions.gpu[:,
:total_num_scheduled_tokens].copy_(...)` triggers a CPU-NPU
synchronization, which blocks subsequent operator dispatch and causes
bubbles visible in Profiling.
This PR changes to full tensor copy
(`mrope_positions.gpu.copy_(mrope_positions.cpu)`) to eliminate the sync
point. The trade-off is a negligible increase in memory usage since
`mrope_positions.cpu` is a small tensor.
**Result:** ~2-3% TPOT improvement with the profiling bubbles
eliminated.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Verified via Profiling that the CPU sync bubble is eliminated and TPOT
is reduced by 2-3%.
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
Signed-off-by: SlightwindSec <slightwindsec@gmail.com>
Co-authored-by: wanghuanjun2113 <wanghuanjun2113@gmail.com>
### What this PR does / why we need it?
The community has added a cleaning mechanism for the metadata after the
main model finishes running. The MTP layer should not clean the
metadata, and a new condition has been added to avoid cleaning it.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
By ci
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: wangxiaoteng <wangxiaoteng@huawei.com>
### What this PR does / why we need it?
If some `eagle3` model without embed_tokens works with `quarot` target
model, the acceptence rate will drop.
We solve it in this PR.
The relative vllm pr is https://github.com/vllm-project/vllm/pull/36225.
- vLLM main:
4034c3d32e
Signed-off-by: drslark <slarksblood@qq.com>
### What this PR does / why we need it?
When eagle and cp are enabled at the same time, there is an error in
pcp_allgather due to hidden_states. This PR fixes this issue.
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
---------
Signed-off-by: lilinsiman <lilinsiman@gmail.com>
### What this PR does / why we need it?
Support FlashComm1 for Qwen3-Next. Fix some padding problems in Sequence
Parallel (SP)
and resolve precision problems in shared_out when both FlashComm1 is
enabled.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
CI
- vLLM version: v0.15.0
- vLLM main:
83b47f67b1
---------
Signed-off-by: zhaojiangjiang <zhaojiangjiang1@h-partners.com>
Co-authored-by: zhaojiangjiang <zhaojiangjiang1@h-partners.com>
### 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>
## Problem
When MTP is enabled, prefill requests with `prompt_tokens ==
num_spec_tokens + 1` are incorrectly classified as decode requests,
causing accuracy issues.
## Root Cause
The `uniform_decode` condition only checked:
- `max_num_scheduled_tokens == uniform_decode_query_len`
- `num_tokens == max_num_scheduled_tokens * num_reqs`
This is insufficient because a prefill request with specific prompt
length satisfies these conditions as well.
## Fix
Add `is_all_decode` check to ensure all requests have
`num_computed_tokens > 0` before classifying as uniform decode, since
decode requests must have computed at least one token.
- vLLM version: v0.15.0
- vLLM main:
83b47f67b1
---------
Signed-off-by: SlightwindSec <slightwindsec@gmail.com>
### What this PR does / why we need it?
To analyze the overhead of the dynamic eplb adjustment framework in
detail, we added the time consumption of the adjustment to the print
information in profiling mode.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
Signed-off-by: shenchuxiaofugui <1311027364@qq.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?
When using the target model after rotational quantization, the
acceptance rate decreases because the fc weight of the draft model has
not undergone rotational quantization(issue: #6445). We fixed this issue
by performing rotation quantization on the fc weight of the draft model
in the same way as the main model when loading draft model.
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
Signed-off-by: zhaomingyu <zhaomingyu13@h-partners.com>
### What this PR does / why we need it?
ds3.2 pcp supports the combination of MTP and chunkprefill features.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
---------
Signed-off-by: weiguihua2 <weiguihua2@huawei.com>
### What this PR does / why we need it?
PR #5632 introduced a bug by replacing some branches gated by enable_sp
with enable_flash_comm_v1. As a result, when enable_shared_expert_dp is
enabled alone (i.e., VLLM_ASCEND_ENABLE_FLASHCOMM1=0 and
VLLM_ASCEND_ENABLE_FLASHCOMM=0), the behavior becomes inconsistent with
the previous logic and leads to accuracy issues. This PR restores the
original enable_sp-based branching to recover expected behavior and
accuracy.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
#### 1. start server
``` bash
vllm serve /home/weights/DeepSeek-V2-Lite-W8A8/ \
--port 8001 \
--served-model-name auto \
--max-model-len 1024 \
--enforce-eager \
--tensor-parallel-size 2 \
--data-parallel-size 2 \
--gpu-memory-utilization 0.9 \
--enable-expert-parallel \
--additional-config '{"enable_shared_expert_dp": true}'
```
#### 2. curl
```bash
curl -s http://localhost:8001/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "auto",
"messages": [
{"role": "user", "content": "Hello. I have a question. Who are you?"}
],
"max_tokens": 10,
"temperature": 0.0,
"ignore_eos_token": true
}'
```
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
Signed-off-by: realliujiaxu <realliujiaxu@163.com>
### What this PR does / why we need it?
This was meant to be merged in #6536, but I accidentally restored a
commit. You can find the relevant discussion
[here](https://github.com/vllm-project/vllm-ascend/pull/6536#issuecomment-3882883471).
Since `self.pass_config.enable_sp` is forcibly set to `False` in the
[source
code](f176443446/vllm/config/compilation.py (L1066)),
this section will no longer verify whether the generated cudagraph
shapes are multiples of both `uniform_decode_query_len`
(`num_speculative_tokens + 1`) and `tensor_parallel_size`.
This PR enables the `num_speculative_tokens + 1` and
`tensor_parallel_size` check upfront. Therefore, it won't silently round
up the `cudagraph_size` and throw a cryptic error for the user.
A typical example of this cryptic error looks like:
```
ValueError: could not broadcast input array from shape (196,) into shape (14,)
```
### Does this PR introduce _any_ user-facing change?
no.
### How was this patch tested?
Have passed all test.
- vLLM version: v0.15.0
- vLLM main:
83b47f67b1
---------
Signed-off-by: lilinsiman <lilinsiman@gmail.com>
Signed-off-by: guozr <guozr1997@hotmail.com>
Co-authored-by: lilinsiman <lilinsiman@gmail.com>
Co-authored-by: drslark <slarksblood@qq.com>
Co-authored-by: guozr <guozr1997@hotmail.com>
### What this PR does / why we need it?
There will be random ouputs if we run model with GDN attention in graph
mode:
```python
prompts = [
"1. Who are you?",
]
sampling_params = SamplingParams(temperature=0.6, top_p=0.95, top_k=40, max_tokens=32)
sampling_params = SamplingParams(temperature=0.0, top_p=0.95, top_k=40, max_tokens=5)
llm = LLM(model="/home/model/Qwen3-Next-80B-A3B-Instruct",
tensor_parallel_size=4,
distributed_executor_backend="mp",
gpu_memory_utilization=0.7,
speculative_config={
"method": "qwen3_next_mtp",
"num_speculative_tokens": 3,
},
compilation_config={
"cudagraph_mode": "FULL_DECODE_ONLY",
"cudagraph_capture_sizes": [8],
},
max_model_len=4096,
enable_prefix_caching=False)
outputs = llm.generate(prompts, sampling_params)
for output in outputs:
prompt = output.prompt
generated_text = output.outputs[0].text
print(f"{output.prompt_token_ids=}")
print(f"{output.outputs[0].token_ids=}")
print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
```
Before appling this change, the outputs was:
```text
output.prompt_token_ids=[16, 13, 10479, 525, 498, 30]
output.outputs[0].token_ids=[3555, 323, 279, 1112, 279]
Prompt: '1. Who are you?', Generated text: ' What and the... the'
```
After applying this change, the output is:
```text
output.prompt_token_ids=[16, 13, 10479, 525, 498, 30]
output.outputs[0].token_ids=[3555, 374, 697, 829, 30]
Prompt: '1. Who are you?', Generated text: ' What is your name?'
```
**Why does this change sovle the problem?**
Now, `query_start_loc` is padded because of `fia`.
But, for `gdn-attention`, padded version of `query_start_loc` will cause
accuracy problem.
So, we need an unpadded version of `query_start_loc` named
`gdn_query_start_loc` and use it in `gdn-attention`, it works fine.
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
As described aboved.
- vLLM version: v0.15.0
- vLLM main:
83b47f67b1
Signed-off-by: drslark <slarksblood@qq.com>
### What this PR does / why we need it?
[Refactor][EAGLE] 7/N Merged PCP and disable_padded interface into
eagle_proposer.py
This pull request significantly refactors the speculative decoding
mechanism by merging Parallel Context Processing (PCP) and Multi-Token
Prediction (MTP) functionalities directly into the eagle_proposer.py.
The changes aim to enhance the efficiency and correctness of distributed
speculative decoding, particularly by enabling the Eagle feature to work
seamlessly with the disable_padded interface. This involves detailed
adjustments to attention metadata, input/output processing, and state
management to ensure proper operation in parallel environments.
1. The PCP and MTP features are migrated to the eagle_proposer.py
2. The Eagle and PCP features are integrated
3. Enable the eagle feature to use the disable_padded interface
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Tests and UT
- vLLM version: v0.15.0
- vLLM main:
83b47f67b1
---------
Signed-off-by: lilinsiman <lilinsiman@gmail.com>
### What this PR does / why we need it?
fix(mtp): resolve MTP core bugs and enhance eager mode test cases
1. Resolved critical issues in eager mode MTP core execution logic;
2. Fixed functional bugs in the _update_states_after_model_execute
function;
3. Updated and released test_mtp_qwen3_next.py to validate eager mode
acceptance rate.
### Does this PR introduce _any_ user-facing change?
None
### How was this patch tested?
- vLLM version: v0.15.0
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.15.0
Signed-off-by: Bowen-Leee <caoshankuangren@gmail.com>
### What this PR does / why we need it?
This PR fixes a bug in `vllm_ascend/worker/model_runner_v1.py` where the
`@torch.inference_mode` decorator was used without parentheses. Using
the decorator without instantiation is deprecated and may not correctly
disable gradient calculations, leading to performance degradation and
increased memory usage during inference. This change adds the required
parentheses to ensure `torch.inference_mode` is applied correctly.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
The change is a minor syntax correction. Existing CI tests should cover
this.
- vLLM version: v0.15.0
- vLLM main:
9562912cea
Signed-off-by: gcanlin <canlinguosdu@gmail.com>
### What this PR does / why we need it?
1. Following https://github.com/vllm-project/vllm/pull/32322, use the
`memory_profiling` context manager from vllm for profiling.
2. Fix wrong non-torch memory value recorded during profiling, which is
not its peak during inference.
---
**More details about point 2:**
After profling, the non-torch memory value we recorded is lower than
that in real inference. This is mainly because of the different memory
management behaviour between `torch.cuda.empty_cache()` and
`torch.npu.empty_cache()`.
With regard to `torch.cuda.empty_cache()`, it only recycle the unused
memory in pytorch memory pool (i.e., memory managed by pytorch caching
allocator), **with no affect to non-torch memory**. However, as for
`torch.npu.empty_cache()`, it has a totally different memory management
mechanism, i.e., it may call `aclrtSynchronize` and **enable Ascend
runtime to free up non-torch memory**.
Thus, the non-torch memory value we recorded after
`torch.npu.empty_cache()` is much lower than its peak during profling.
Resolution:
We record the peak non-torch memory value
(`non_torch_memory_before_empty_cache`) after profiling, but before
`torch.npu.empty_cache()`. Then, we add the diff
(`non_torch_memory_cleared_by_empty_cache =
non_torch_memory_before_empty_cache - self.non_torch_memory`) to
non-torch memory when calculating available KV cache memory, which will
lead to less KV cache memory (i.e., it's safer to avoid OOM issues).
---
> [!NOTE]
> This PR needs to wait for main2main aligning to latest vllm commit
before merging.
### Does this PR introduce _any_ user-facing change?
no.
### How was this patch tested?
Before this PR, the non-torch memory we used to calculate available KV
cache memory is **0.90 G**, whereas its peak during real inference is
**1.08 G**, diff: **182.00 M**.
After this PR, we add this diff to non-torch memory after profiling and
thus make the profiling results more accurate.
- vLLM version: v0.15.0
- vLLM main:
d7e17aaacd
---------
Signed-off-by: shen-shanshan <467638484@qq.com>
### What this PR does / why we need it?
#6043 deleted the forward_before phase of the dynamic eplb. Currently,
the end-to-end precision is monitored in the UT, and the log is not
printed in the key place. As a result, the eplb does not take effect and
is not intercepted.
1. The forward_before function is added back.
2. Delete unnecessary logs and add key logs.
3. Warm-up of algorithm 3 is added.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
#### The conversation is normal.
Okay, the user is asking, \"What is deep learning?\" I need to explain
this in a clear and concise way. Let me start by recalling what I know
about deep learning. It's a subset of machine learning, right? So first,
I should mention that it's part of machine learning, which itself is a
branch of AI. Then, the key aspect of deep learning is the use of neural
networks with multiple layers. These are called deep neural
networks.\n\nWait, I should define neural networks first. Maybe start
with the basics. A neural network is inspired by the human brain, with
layers of nodes (neurons) that process data. But deep learning
specifically refers to networks with many layers—hence \"deep.\" So the
term \"deep\" comes from the number of layers. \n\nI should explain how
deep learning works. It involves training these networks on large
datasets, allowing them to automatically learn features from the data.
Unlike traditional machine learning, where you might have to manually
extract features, deep learning models can do this automatically. That's
a key point. For example, in image recognition, a deep learning model
can learn to detect edges, shapes, and then more complex patterns
without human intervention.\n\nApplications are important too. The user
might want to know where deep learning is used. Common examples include
image and speech recognition, natural language processing, autonomous
vehicles, and recommendation systems. Maybe mention specific
technologies like self-driving cars using computer vision or virtual
assistants like Siri or Alexa
- vLLM version: v0.15.0
- vLLM main:
13397841ab
Signed-off-by: shenchuxiaofugui <1311027364@qq.com>
### What this PR does / why we need it?
The incorrect regular expression syntax `.*[UE4M3|ue4m3].*` actually
ignores all words containing any of the following characters: `u, e, 4,
m, 3, |`
```yaml
extend-ignore-identifiers-re = [".*Unc.*", ".*_thw",
".*UE8M0.*", ".*[UE4M3|ue4m3].*", ".*eles.*", ".*fo.*", ".*ba.*",
".*ot.*", ".*[Tt]h[rR].*"]
```
===fix===>
```yaml
extend-ignore-identifiers-re = [".*Unc.*", ".*_thw",
".*UE8M0.*", ".*(UE4M3|ue4m3]).*", ".*eles.*", ".*fo.*", ".*ba.*",
".*ot.*", ".*[Tt]h[rR].*"]
```
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.15.0
- vLLM main:
9562912cea
Signed-off-by: MrZ20 <2609716663@qq.com>
### What this PR does / why we need it?
This PR extends original `rope_triton_forward` and
`split_qkv_rmsnorm_rope` to support `cos_sin_cache` && `positions` as
inputs. This fully aligns to vLLM RoPE api interface. Compared with
earlier implementation for RoPE, the benefits are:
1. avoiding pre-computation of `cos` `sin` before model execution, which
helps to remove redundant codes.
2. allowing eagle3 draft model to have different rope parameters with
main model (see #6612 ). This help to recover accept rate && accuracy in
that case.
In addition, this kernel change only introduces very small performance
degradation. Those `index_select` or `chunk` operations are now changed
into simple memory access in triton kernel (For example,
https://github.com/vllm-project/vllm-ascend/pull/5450/changes#diff-a4c2d3071530df193b98f9bf38553874bc4d47571336711f116c26d019cfbb6aR77-R81).
**Highlights**
- **RoPE Cache Unification**: Replaced separate _sin and _cos global
tensors with a unified cos_sin_cache and explicit positions tensor for
Rotary Positional Embeddings (RoPE), streamlining data handling.
- **Triton Kernel Integration**: Updated Triton kernels
(split_qkv_rmsnorm_rope_kernel, _triton_rope) to directly consume the
cos_sin_cache and positions for more efficient and integrated RoPE
calculations.
- **Custom Operation Registration**: Registered `rope_forward_oot` as a
new custom operation, allowing its use in fused compilation passes and
providing a dedicated entry point for the new RoPE implementation.
- **Refactored RoPE Forward Pass**: Modified the rope_forward_oot
function to accept the new cos_sin_cache and positions arguments,
enabling a more flexible and integrated RoPE application within the
system.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
5326c89803
Additional test on Qwen3-235b accuracy:
| Aime2024 | GSM8K | Livecodebench |
| -------- | -------- | -------- |
| 83.33 | 96.26 | 70.23 |
---------
Signed-off-by: Angazenn <supperccell@163.com>
### What this PR does / why we need it?
1. Currently, eplb registers different attributes for different models,
but these attributes are not actually used. Now, these attributes are
directly deleted.
2. Add some log about eplb.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
#### Deepseek v3.1 chat
Of course! Here is a comprehensive explanation of deep learning, broken
down for clarity.\n\n### The Simple Analogy: A Child Learning to
Recognize a Cat\n\nImagine teaching a child what a cat is. You don't
give them a rulebook with instructions like \"has pointy ears, whiskers,
and a tail.\" Instead, you show them many pictures, saying \"this is a
cat\" or \"this is not a cat.\" The child's brain gradually learns to
identify the complex patterns—the combination of shapes, colors, and
textures—that define \"cat-ness.\"\n\n**Deep learning is essentially
this, but for computers.** It's a method for teaching computers to learn
from examples and recognize patterns directly from data (like images,
sound, or text) without being explicitly programmed with rigid
rules.\n\n---\n\n### The Technical Definition\n\n**Deep Learning is a
subfield of machine learning, which itself is a subfield of artificial
intelligence (AI).** It uses artificial **neural networks** with many
layers (\"deep\" networks) to model and understand complex patterns in
data.\n\nHere are the key concepts in that definition:\n\n1.
**Artificial Intelligence (AI):** The broad science of making machines
smart and capable of performing tasks that typically require human
intelligence.\n2. **Machine Learning (ML):** A subset of AI that gives
computers the ability to learn from data *without* being explicitly
programmed for every single rule.\n3. **Deep Learning (DL):** A
specific, powerful
- vLLM version: v0.15.0
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.15.0
Signed-off-by: shenchuxiaofugui <1311027364@qq.com>
