backport of #7474
This PR adds C8 (INT8) KV cache quantization support for standard GQA
attention models (e.g., Qwen3-32B W8A8C8). C8 uses static per-channel
quantization scales to store KV cache in INT8, reducing KV cache memory
by ~50% compared to BF16, enabling higher batch concurrency and longer
context lengths on the same hardware.
**Key changes:**
1. **`attention_v1.py`** — New `AscendC8AttentionBackendImpl` subclass
of `AscendAttentionBackendImpl`:
- `_prepare_c8_scales`: Shards per-channel scales/offsets to the current
TP rank and pre-computes BF16 BNSD-shaped antiquant tensors (one-time
per layer).
- `_quantize_kv_to_int8`: Quantizes BF16 K/V to INT8 before
`reshape_and_cache`, using pre-cached inverse scales.
- `_forward_c8_decode`: FIA V1 BNSD paged attention with native INT8 KV
and `perchannel` antiquant mode.
- `_forward_c8_chunked_prefill`: Splits decode (FIA V1 BNSD paged INT8)
and prefill (FIA V1 TND float) into two kernel calls.
- `_forward_c8_fused_infer_attention`: Handles `PrefillNoCache` and
`PrefillCacheHit` states.
2. **`quantization/methods/kv_c8.py`** — New
`AscendC8KVCacheAttentionMethod` scheme:
- Creates `k/v_cache_scale/offset` parameters via
`_c8_kv_scale_weight_loader`, which handles per-channel scale shapes and
lazy resizing.
- Sets `layer.kv_cache_torch_dtype = torch.int8` so
`get_kv_cache_spec()` returns INT8 dtype automatically.
- Upgrades `layer.impl` to `AscendC8AttentionBackendImpl` via class
surgery.
3. **`quantization/modelslim_config.py`** — C8 branch in
`get_quant_method()` activates when `kv_cache_type == "C8"` in
`quant_model_description.json`.
4. **`patch/worker/patch_qwen3_c8.py`** — Intercepts per-channel C8
scale/offset weights before `AutoWeightsLoader` discards them, routing
them to the parameters created by `AscendC8KVCacheAttentionMethod`.
5. **`tests/ut/quantization/test_kv_c8.py`** — Unit tests covering
`_c8_kv_scale_weight_loader`, `AscendC8KVCacheAttentionMethod`, and
`AscendC8AttentionBackendImpl` scale helpers.
Yes. Users can now serve Qwen3-32B W8A8C8 quantized models with INT8 KV
cache on Ascend NPU. The model checkpoint must contain a
`quant_model_description.json` with `"kv_cache_type": "C8"` and
per-channel scale/offset tensors in safetensors.
No changes to the serving CLI — the feature activates automatically when
the quantization config is detected.
Benchmarked with `vllm serve` (TP=8, `max_num_seqs=256`,
`max_model_len=131072`, `enable_chunked_prefill=true`) + `random_bench`
(input_len=10240, output_len=2048, 960 prompts, max_concurrency=192):
```
============ Serving Benchmark Result ============
Successful requests: 960
Failed requests: 0
Maximum request concurrency: 192
Benchmark duration (s): 1359.81
Total input tokens: 9830400
Total generated tokens: 1966080
Request throughput (req/s): 0.71
Output token throughput (tok/s): 1445.85
Peak output token throughput (tok/s): 2304.00
Total token throughput (tok/s): 8675.12
---------------Time to First Token----------------
Mean TTFT (ms): 24598.51
Median TTFT (ms): 23167.02
P50 TTFT (ms): 23167.02
P90 TTFT (ms): 47717.08
P99 TTFT (ms): 84402.61
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms): 120.76
Median TPOT (ms): 121.50
P50 TPOT (ms): 121.50
P90 TPOT (ms): 127.05
P99 TPOT (ms): 130.13
---------------Inter-token Latency----------------
Mean ITL (ms): 120.70
Median ITL (ms): 90.34
P50 ITL (ms): 90.34
P90 ITL (ms): 93.79
P99 ITL (ms): 101.80
==================================================
```
All attention states verified: `PrefillNoCache`, `PrefillCacheHit`,
`ChunkedPrefill`, `DecodeOnly`.
- vLLM version: v0.17.0
- vLLM main:
8b6325758c
Signed-off-by: lico67373 <918688502@qq.com>
Co-authored-by: LICO67373 <110013619+LICO1314@users.noreply.github.com>
### What this PR does / why we need it?
This PR simplifies the apply method in w8a8_dynamic.py by removing the
conditional logic that used fused_w1_scale and fused_w2_scale based on
the fused_scale_flag. This redundant wrap behavior leads to EPLB break
in int8 quantization scenarios.
Cherry-picked from #7188. Note that only bugfix lines in that PR are
picked.
Signed-off-by: linfeng-yuan <1102311262@qq.com>
### What this PR does / why we need it?
This PR fixes A5 MXFP8 MoE scale handling in the fused MoE path.
- It normalizes MXFP8 activation scales to the packed 3D layout expected
by A5 kernels, including both precomputed dynamic_scale inputs and gmm1
output scales before they are consumed by downstream grouped matmul ops.
- It also refines the MXFP8 force load-balancing path in profiling runs.
- This PR also enables npu_gating_top_k from torch_npu instead of custom
op when running ascend950 chip.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
CI and E2E serving tests on Ascend950DT passed.
---------
Signed-off-by: linfeng-yuan <1102311262@qq.com>
### What this PR does / why we need it?
During the attention quantization process of DeepSeek V3.2, it is
necessary to retrieve the Hadamard matrix from the weights to facilitate
the computation.
### Does this PR introduce _any_ user-facing change?
No. But there will be two new tensor in quant weight.
### How was this patch tested?
- vLLM version: v0.18.0
- vLLM main:
8b6325758c
---------
Signed-off-by: mayumeng <m30059191@china.huawei.com>
Co-authored-by: mayumeng <m30059191@china.huawei.com>
### What this PR does / why we need it?
Refactor `vllm_ascend/ops/fused_moe` to replace scattered MoE business
`**kwargs` with typed request objects and explicit stage boundaries.
- Prepare, dispatch, MLP, and quant stages now have clearer ownership.
- Main MoE path no longer depends on business `kwargs.get(...)` lookups.
- Comm and dispatcher interfaces are request-only on the main path.
- UTs can assert stage-level fields directly instead of inferring
behavior indirectly.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
CI passed.
---------
Signed-off-by: linfeng-yuan <1102311262@qq.com>
### What this PR does / why we need it?
This PR adapts the `w8a8_mxfp8` quantization method to support Qwen
Vision-Language (VL) models. Key changes include:
- Reshaping multi-dimensional input tensors to 2D before the quantized
matrix multiplication.
- Reshaping the 2D output back to its original multi-dimensional format.
- Adding specific output reshaping for the visual components of Qwen VL
models.
- Casting the bias tensor to `float32` to comply with the
`npu_quant_matmul` kernel requirements.
These changes are necessary to enable `w8a8_mxfp8` quantization for
models with multi-modal inputs like Qwen VL.
### Does this PR introduce _any_ user-facing change?
No, this is a backend enhancement to extend quantization support to new
model architectures. There are no user-facing API or behavior changes.
### How was this patch tested?
CI is expected to pass. Manual testing should be performed with a Qwen
VL model using `w8a8_mxfp8` quantization to verify correctness and
performance.
- vLLM version: v0.17.0
- vLLM main:
4497431df6
---------
Signed-off-by: ksiyuan <ksiyuan@umich.edu>
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?
This PR aims to support aclgraph for model runner v2, please see RFC
#5208. The PR contains these modifications:
- adapt to newest commit of vllm main branch.
- supply a unified interface of extra forward context for both model
runner v1 and model runner v2.
- implement graph mode for main model.
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
### What this PR does / why we need it?
This pull request addresses a bug related to the fused mc2 functionality
within the EPLB (Expert Parallelism Load Balancing) system, specifically
impacting quantization and MoE communication.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.15.0
- vLLM main:
83b47f67b1
Signed-off-by: Spicy-Stick <873805887@qq.com>
Signed-off-by: root <root@localhost.localdomain>
### What this PR does / why we need it?
This pull request is for quantization adaptation of Qwen3Omni, and it
achieves operator-level optimization and AUT (Auto-Quantization Tuning)
component optimization through patch-based modifications.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.15.0
- vLLM main:
83b47f67b1
---------
Signed-off-by: tanhaoan333 <tanhaoan@huawei.com>
### What this PR does / why we need it?
`mxfp_compat` only provides dtype/symbol compatibility helpers for
different `torch_npu` versions, but it was placed under
`vllm_ascend.quantization`. Importing it from device/ops paths could
trigger `quantization/__init__.py` and pull in heavy quantization method
dependencies, increasing startup coupling and causing import-cycle risk
(especially on 310P paths).
### Does this PR introduce _any_ user-facing change?
No functional behavior change intended.
### How was this patch tested?
CI passed.
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
---------
Signed-off-by: linfeng-yuan <1102311262@qq.com>
## Summary
- Remove unused `set_rotation_config` and `apply_rotation` methods from
`AscendW4A4LaosDynamicLinearMethod`
- Remove unused `rotation_type` field and associated conditional
quantization parameters (`heads_rotation`, `kronecker_rotation_n`,
`kronecker_rotation_m`)
These rotation-related functions and parameters are never called in the
current W4A4 LAOS dynamic quantization workflow.
- vLLM version: v0.15.0
- vLLM main:
d7e17aaacd
Signed-off-by: SlightwindSec <slightwindsec@gmail.com>
### What this PR does / why we need it?
Introduced 310P W8A8 Quantization Support: New modules and methods have
been added to enable W8A8 static quantization specifically for the
Ascend 310P platform.
Platform-Specific Quantization Configuration Loading: The system now
dynamically loads the appropriate quantization configurations
(AscendCompressedTensorsConfig, AscendModelSlimConfig) based on whether
the current hardware is an Ascend 310P device.
Implemented AscendW8A8LinearMethod310P: A dedicated linear quantization
method for 310P is provided, handling the specifics of weight and
activation quantization, including input parameter broadcasting and
weight data manipulation.
Extended AscendModelSlimConfig for 310P: A specialized configuration
class for 310P integrates the new W8A8 linear method for both standard
linear layers and vocabulary parallel embeddings, ensuring proper
quantization application.
- vLLM version: v0.14.1
- vLLM main:
dc917cceb8
---------
Signed-off-by: Tflowers-0129 <2906339855@qq.com>
Signed-off-by: Shaoxu Cheng <2906339855@qq.com>
### What this PR does / why we need it?
While using the LLM Compressor quantization tool from the VLLM community
to generate quantized weights, the VLLM Ascend engine needs to be
adapted to support the compressed tensors quantization format.
1. Support Moe model W4A8 dynamic weight.
- vLLM version: v0.13.0
- vLLM main:
bde38c11df
---------
Signed-off-by: LHXuuu <scut_xlh@163.com>
Signed-off-by: menogrey <1299267905@qq.com>
Co-authored-by: menogrey <1299267905@qq.com>
