EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[quantization] Add w8a16 quantization support (#4541)

Browse Source 
### What this PR does / why we need it?
related to https://github.com/vllm-project/vllm-ascend/issues/4267

### Does this PR introduce _any_ user-facing change?
support w8a16 quantization now

### How was this patch tested?

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

### Test
tested using [aisbench](https://gitee.com/aisbench/benchmark/) with tp2
#### Precision
  | ceval | mmlu | gsm8k
-- | -- | -- | --
bf16 | 90.46 | 89.17 | 96.21
w8a16 | 89.51 | 89.29 | 95.98

#### Performance
  | input_len | output_len | concurrency | TTFT (ms) | TPOT (ms) | TPS
(Total) (tokens/s)
-- | -- | -- | -- | -- | -- | --
bf16 | 2048 | 2048 | 10 | 1911.7136 | 77.988 | 253.9866
w8a16 | 2048 | 2048 | 10 | 2128.6334 | 67.1633 | 293.9117
bf16 | 3500 | 1024 | 10 | 3076.2509 | 84.3525 | 506.949
w8a16 | 3500 | 1024 | 10 | 2685.2031 | 73.015 | 585.4717

---------

Signed-off-by: yyt <yangyit139@gmail.com>
Signed-off-by: TmacAaron <yangyit139@gmail.com>
Co-authored-by: realliujiaxu <realliujiaxu@163.com>
This commit is contained in:
TmacAaron
2025-12-24 19:49:32 +08:00
committed by GitHub
parent 515267de22
commit 5018f2d8fd
4 changed files with 229 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
tests/e2e/singlecard/test_quantization.py
View File

@@ -18,6 +18,7 @@
from modelscope import snapshot_download # type: ignore[import-untyped]
from tests.e2e.conftest import VllmRunner
from tests.e2e.model_utils import check_outputs_equal
def test_qwen3_w8a8_quant():
@@ -25,10 +26,53 @@ def test_qwen3_w8a8_quant():
example_prompts = [
"vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs."
]
vllm_target_outputs = [([
85, 4086, 44, 374, 264, 1550, 42747, 628, 323, 4938, 72816, 44378, 323,
13480, 4712, 369, 444, 10994, 82, 13, 1084, 374, 6188, 311, 387
], 'vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. It is designed to be'
)]
with VllmRunner(
snapshot_download("vllm-ascend/Qwen3-0.6B-W8A8"),
max_model_len=8192,
gpu_memory_utilization=0.7,
quantization="ascend",
) as vllm_model:
vllm_model.generate_greedy(example_prompts, max_tokens)
vllm_quant_w8a8_outputs = vllm_model.generate_greedy(
example_prompts, max_tokens)
check_outputs_equal(
outputs_0_lst=vllm_target_outputs,
outputs_1_lst=vllm_quant_w8a8_outputs,
name_0="vllm_target_outputs",
name_1="vllm_w8a16_outputs",
)
def test_qwen3_dense_w8a16():
max_tokens = 5
example_prompts = [
"vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs."
]
vllm_target_outputs = [([
85, 4086, 44, 374, 264, 1550, 42747, 628, 323, 4938, 72816, 44378, 323,
13480, 4712, 369, 444, 10994, 82, 13, 1084, 374, 6188, 311, 387
], 'vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. It is designed to be'
)]
with VllmRunner(
snapshot_download("vllm-ascend/Qwen3-0.6B-W8A16"),
max_model_len=8192,
enforce_eager=False,
gpu_memory_utilization=0.7,
quantization="ascend",
) as vllm_model:
vllm_quant_w8a16_outputs = vllm_model.generate_greedy(
example_prompts, max_tokens)
check_outputs_equal(
outputs_0_lst=vllm_target_outputs,
outputs_1_lst=vllm_quant_w8a16_outputs,
name_0="vllm_target_outputs",
name_1="vllm_w8a16_outputs",
)