minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity
Files
3abdbab9bbb4ef960be30809a62affd6c6d7a38f
sglang/docs_new/docs/hardware-platforms/ascend-npus/ascend_npu_quantization.mdx
T

508 lines
19 KiB
Plaintext
Raw Blame History

---
title: "Quantization on Ascend"
metatags:
description: "Load, export, and serve quantized models on Ascend NPUs with SGLang."
---
To load already quantized models, simply load the model weights and config. Again, if the model has been quantized offline, there's no need to add `--quantization` argument when starting the engine. The quantization method will be automatically parsed from the downloaded `quant_model_description.json` or `config.json` config.
SGLang supports **mix-bits** quantization (independently defines and loads each layer depending on the type of quantification specified in the `quant_model_description.json`). [Advanced mix-bits for MoE](https://github.com/sgl-project/sglang/pull/17361) in progress, will add independent quantization determination for the w13 (up-gate) and w2 (down) layers.
[ModelSlim on Ascend support](https://github.com/sgl-project/sglang/pull/14504)
<table>
<thead>
<tr>
<th>Quantization scheme</th>
<th>Layer type</th>
<th>A2 Supported</th>
<th>A3 Supported</th>
<th>A5 Supported</th>
<th>Diffusion models</th>
</tr>
</thead>
<tbody>
<tr>
<td>W4A4 dynamic</td>
<td>Linear</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
</tr>
<tr>
<td>W8A8 static</td>
<td>Linear</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
</tr>
<tr>
<td>W8A8 dynamic</td>
<td>Linear</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
</tr>
<tr>
<td>MXFP8 (<a href="https://github.com/sgl-project/sglang/pull/20922">Diffusion</a>, <a href="https://github.com/sgl-project/sglang/pull/22352">LLM dense</a>)</td>
<td>Linear</td>
<td><strong style={{color: 'red'}}>x</strong></td>
<td><strong style={{color: 'red'}}>x</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
</tr>
<tr>
<td><a href="https://github.com/sgl-project/sglang/pull/22338">MXFP4</a></td>
<td>Linear</td>
<td><strong style={{color: 'red'}}>x</strong></td>
<td><strong style={{color: 'red'}}>x</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
</tr>
<tr>
<td><a href="https://github.com/sgl-project/sglang/pull/23650">MXFP4 W4A8</a></td>
<td>Linear</td>
<td><strong style={{color: 'red'}}>x</strong></td>
<td><strong style={{color: 'red'}}>x</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'red'}}>x</strong></td>
</tr>
<tr>
<td>W4A4 dynamic</td>
<td>MoE</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
<td><strong style={{color: 'red'}}>x</strong></td>
</tr>
<tr>
<td>W4A8 dynamic</td>
<td>MoE</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
<td><strong style={{color: 'red'}}>x</strong></td>
</tr>
<tr>
<td>W8A8 dynamic</td>
<td>MoE</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
<td><strong style={{color: 'red'}}>x</strong></td>
</tr>
<tr>
<td><a href="https://github.com/sgl-project/sglang/pull/20922">MXFP8</a></td>
<td>MoE</td>
<td><strong style={{color: 'red'}}>x</strong></td>
<td><strong style={{color: 'red'}}>x</strong></td>
<td><strong style={{color: 'blue'}}>WIP</strong></td>
<td><strong style={{color: 'red'}}>x</strong></td>
</tr>
</tbody>
</table>
[AWQ on Ascend support](https://github.com/sgl-project/sglang/pull/10158):
<table>
<thead>
<tr>
<th>Quantization scheme</th>
<th>Layer type</th>
<th>A2 Supported</th>
<th>A3 Supported</th>
<th>A5 Supported</th>
</tr>
</thead>
<tbody>
<tr>
<td>W4A16</td>
<td>Linear</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
</tr>
<tr>
<td>W8A16</td>
<td>Linear</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
</tr>
<tr>
<td>W4A16</td>
<td>MoE</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
</tr>
</tbody>
</table>
GPTQ on Ascend support
<table>
<thead>
<tr>
<th>Quantization scheme</th>
<th>Layer type</th>
<th>A2 Supported</th>
<th>A3 Supported</th>
<th>A5 Supported</th>
</tr>
</thead>
<tbody>
<tr>
<td><a href="https://github.com/sgl-project/sglang/pull/15203">W4A16</a></td>
<td>Linear</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
</tr>
<tr>
<td><a href="https://github.com/sgl-project/sglang/pull/15203">W8A16</a></td>
<td>Linear</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
</tr>
<tr>
<td><a href="https://github.com/sgl-project/sglang/pull/16364">W4A16 MOE</a></td>
<td>MoE</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
</tr>
<tr>
<td><a href="https://github.com/sgl-project/sglang/pull/16364">W8A16 MOE</a></td>
<td>MoE</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
</tr>
</tbody>
</table>
[Auto-round on Ascend support](https://github.com/sgl-project/sglang/pull/16699)
<table>
<thead>
<tr>
<th>Quantization scheme</th>
<th>Layer type</th>
<th>A2 Supported</th>
<th>A3 Supported</th>
<th>A5 Supported</th>
</tr>
</thead>
<tbody>
<tr>
<td>W4A16</td>
<td>Linear</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
</tr>
<tr>
<td>W8A16</td>
<td>Linear</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
</tr>
<tr>
<td>W4A16</td>
<td>MoE</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
</tr>
<tr>
<td>W8A16</td>
<td>MoE</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
</tr>
</tbody>
</table>
Compressed-tensors (LLM Compressor) on Ascend support:
<table>
<thead>
<tr>
<th>Quantization scheme</th>
<th>Layer type</th>
<th>A2 Supported</th>
<th>A3 Supported</th>
<th>A5 Supported</th>
</tr>
</thead>
<tbody>
<tr>
<td><a href="https://github.com/sgl-project/sglang/pull/14504">W8A8 dynamic</a></td>
<td>Linear</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
</tr>
<tr>
<td><a href="https://github.com/sgl-project/sglang/pull/14736">W4A8 dynamic with/without activation clip</a></td>
<td>MoE</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
</tr>
<tr>
<td><a href="https://github.com/sgl-project/sglang/pull/12759">W4A16 MOE</a></td>
<td>MoE</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
</tr>
<tr>
<td><a href="https://github.com/sgl-project/sglang/pull/14504">W8A8 dynamic</a></td>
<td>MoE</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
</tr>
</tbody>
</table>
[GGUF on Ascend support](https://github.com/sgl-project/sglang/pull/17883)
<table>
<thead>
<tr>
<th>Quantization type</th>
<th>Layer type</th>
<th>A2 Supported</th>
<th>A3 Supported</th>
<th>A5 Supported</th>
</tr>
</thead>
<tbody>
<tr>
<td>All GGUF types (standard, K-quant)</td>
<td>Linear</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
</tr>
<tr>
<td>All GGUF types (standard, K-quant)</td>
<td>MoE</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
</tr>
</tbody>
</table>
**Usage Examples:**
- Dense model (e.g. Qwen3-14B-Q4_K_M.gguf):
```bash Command
python3 -m sglang.launch_server \
--model-path Qwen3-14B-Q4_K_M.gguf \
--device npu --attention-backend ascend \
--host 0.0.0.0 --port 30000 \
--mem-fraction-static 0.7 --tp-size 2
```
- MoE model (e.g. Qwen3-30B-A3B-Q4_K_M.gguf):
```bash Command
python3 -m sglang.launch_server \
--model-path Qwen3-30B-A3B-Q4_K_M.gguf \
--device npu --attention-backend ascend \
--host 0.0.0.0 --port 30000 \
--mem-fraction-static 0.8 --tp-size 2
```
> **Implementation Notes:**
>
> - GGUF weights are pre-dequantized to FP16/BF16 during model loading on CPU, then transferred to NPU for inference. This trades higher memory usage for faster runtime performance (no per-forward-pass dequantization overhead).
> - MoE layers use `npu_grouped_matmul` and `npu_moe_init_routing` / `npu_moe_finalize_routing` for high-performance expert computation.
> - TP (tensor parallelism) sharding is supported for both dense and MoE GGUF models.
**MXFP8 for LLM dense models (e.g. Qwen3 / Qwen3.5):**
LLM dense W8A8 MXFP8 Linear support on Ascend was added in [PR #22352](https://github.com/sgl-project/sglang/pull/22352). Requires Ascend A5 series or newer (`npu_dynamic_mx_quant` is not available on A2 / A3).
- Online MXFP8 quantization (BF16/FP16 weights → MXFP8 at load time):
```bash Command
python3 -m sglang.launch_server \
--model-path Qwen/Qwen3-8B \
--quantization mxfp8 \
--device npu --attention-backend ascend \
--host 0.0.0.0 --port 30000 \
--mem-fraction-static 0.8 --tp-size 1
```
- Offline MXFP8 quantization (msmodelslim pre-quantized weights, `W8A8_MXFP8` scheme; no `--quantization` flag needed — auto-detected from `quant_model_description.json`):
```bash Command
python3 -m sglang.launch_server \
--model-path /path/to/Qwen3-8B-W8A8-MXFP8 \
--device npu --attention-backend ascend \
--host 0.0.0.0 --port 30000 \
--mem-fraction-static 0.8 --tp-size 1
```
> **Implementation Notes:**
> - Online path: `Fp8Config.get_quant_method()` dispatches to `NPUMXFP8LinearMethod`. Weights are quantized once at load via `npu_dynamic_mx_quant(weight, dst_type=torch_npu.float8_e4m3fn)` and pre-transposed to `[in, out]`; activations are per-token quantized at inference and matmul runs via `npu_quant_matmul(..., group_sizes=[1, 1, 32])` (block_size = 32).
> - Offline path: `ModelSlimMXFP8Scheme` loads `float8_e4m3fn` weights + `float8_e8m0fnu` block scales pre-exported by msmodelslim. Transpose is kept as a non-contiguous view (`.data` assignment) — calling `.contiguous()` would physically reorder the pre-quantized layout and break the block-scale mapping.
> - MoE MXFP8 (FusedMoE/TP) for LLMs is tracked separately and not part of this PR.
**MXFP4 W4A8 for LLM dense models (e.g. Qwen3 / Qwen3.5):**
LLM dense W4A8 (MXFP4 4-bit weights + MXFP8 8-bit activations) Linear support was added in [PR #23650](https://github.com/sgl-project/sglang/pull/23650). Requires Ascend A5 series or newer.
- Online W4A8 quantization (BF16/FP16 weights → MXFP4 at load time):
```bash Command
python3 -m sglang.launch_server \
--model-path Qwen/Qwen3-8B \
--quantization mxfp_w4a8 \
--device npu --attention-backend ascend \
--host 0.0.0.0 --port 30000 \
--mem-fraction-static 0.8 --tp-size 1
```
- Offline W4A8 quantization (msmodelslim pre-quantized weights, `W4A8_MXFP` scheme; no `--quantization` flag needed — auto-detected from `quant_model_description.json`).
> **Implementation Notes:**
> - Weights are packed FP4 (`float4_e2m1fn_x2`, two nibbles per byte) with a UE8M0 per-block shared exponent (block_size = 32); activations are per-token MXFP8. Matmul runs via `npu_quant_matmul(..., x2_dtype=torch_npu.float4_e2m1fn_x2, group_sizes=[0, 0, 32])`.
> - The packed-FP4 dtype passed to the NPU ops (`dst_type` / `x2_dtype` / `input_dtype`) must be resolved from `torch_npu.float4_e2m1fn_x2` (an int enum), not the `torch.float4_e2m1fn_x2` dtype object, which recent op-plugin builds reject.
> - Online and offline share the same kernel path and layout; they differ only in the weight source (RTN at load vs msmodelslim calibration).
## Diffusion Model Quantization on Ascend NPU
SGLang-Diffusion supports MXFP8 online and offline quantization for diffusion models (such as Wan2.2) on Ascend NPUs. MXFP8 requires A5; the ModelSlim W8A8/W4A4 schemes work on A2/A3.
**Requirements for MXFP8:** CANN ≥ 8.0.RC3, Ascend A5
<table>
<thead>
<tr>
<th>Quantization method</th>
<th><code>quant_type</code> in JSON</th>
<th>Scheme class</th>
<th>Mode</th>
<th>A2/A3 Supported</th>
<th>A5 Supported</th>
<th>Trigger</th>
</tr>
</thead>
<tbody>
<tr>
<td>MXFP8 (W8A8)</td>
<td>—</td>
<td><code>MXFP8Config</code></td>
<td>Online</td>
<td><strong style={{color: 'red'}}>x</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><code>--quantization mxfp8</code></td>
</tr>
<tr>
<td>MXFP8 (W8A8)</td>
<td><code>W8A8_MXFP8</code></td>
<td><code>ModelSlimMXFP8Scheme</code></td>
<td>Offline</td>
<td><strong style={{color: 'red'}}>x</strong></td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td>auto-detected from <code>quant_model_description.json</code></td>
</tr>
<tr>
<td>W8A8 static</td>
<td><code>W8A8</code></td>
<td><code>ModelSlimW8A8Int8</code></td>
<td>Offline</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
<td>auto-detected from <code>quant_model_description.json</code></td>
</tr>
<tr>
<td>W8A8 dynamic</td>
<td><code>W8A8_DYNAMIC</code></td>
<td><code>ModelSlimW8A8Int8</code></td>
<td>Offline</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
<td>auto-detected from <code>quant_model_description.json</code></td>
</tr>
<tr>
<td>W4A4 dynamic</td>
<td><code>W4A4_DYNAMIC</code></td>
<td><code>ModelSlimW4A4Int4</code></td>
<td>Offline</td>
<td><strong style={{color: 'green'}}>√</strong></td>
<td><strong style={{color: 'orange'}}>TBD</strong></td>
<td>auto-detected from <code>quant_model_description.json</code></td>
</tr>
</tbody>
</table>
### Online MXFP8 Quantization
Online quantization dynamically quantizes FP16/BF16 weights to MXFP8 at load time using `npu_dynamic_mx_quant` + `npu_quant_matmul` CANN kernels. Pass `--quantization mxfp8` to override auto-detection.
```bash Command
# Start the diffusion server with online MXFP8 quantization
sglang serve \
--model-path Wan-AI/Wan2.2-T2V-A14B-Diffusers \
--quantization mxfp8 \
--num-gpus 4
```
```bash Command
# One-shot generation
sglang generate \
--model-path Wan-AI/Wan2.2-T2V-A14B-Diffusers \
--quantization mxfp8 \
--prompt "a beautiful sunset over the mountains" \
--save-output
```
### Offline MXFP8 Quantization (ModelSlim)
For offline quantization, pre-quantize the model with msModelSlim and load the resulting checkpoint. The quantization scheme is auto-detected from `quant_model_description.json`, so no extra `--quantization` flag is needed.
**Step 1: Quantize with msModelSlim**
```bash Command
msmodelslim quant \
--model_path /path/to/wan2_2_float_weights \
--save_path /path/to/wan2_2_mxfp8_weights \
--device npu \
--model_type Wan2_2 \
--quant_type mxfp8 \
--trust_remote_code True
```
> Note: SGLang does not support quantized embeddings; disable embedding quantization when using msmodelslim.
**Step 2: Convert to Diffusers format**
msModelSlim saves quantized Wan2.2 weights in the original Wan format. Convert to Diffusers format using the provided repack script:
```bash Command
python python/sglang/multimodal_gen/tools/wan_repack.py \
--input-path /path/to/wan2_2_mxfp8_weights \
--output-path /path/to/wan2_2_mxfp8_diffusers
```
Then copy all files from the original Diffusers checkpoint (except the `transformer`/`transformer_2` folders) into the output directory.
**Step 3: Run inference**
```bash Command
sglang generate \
--model-path /path/to/wan2_2_mxfp8_diffusers \
--prompt "a beautiful sunset over the mountains" \
--save-output
```
For pre-quantized checkpoints available on ModelScope, see [modelscope/Eco-Tech](https://modelscope.cn/models/Eco-Tech).
Reference in New Issue View Git Blame Copy Permalink