[Quantization] add humming quantization kernel (#23754)
Co-authored-by: guzekai01 <zekai01@antgroup.com> Co-authored-by: Julian Huang <huangzhilin.hzl@gmail.com> Co-authored-by: 墨楼 <huangzhilin.hzl@antgroup.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: Peng Zhang <aniz1905@gmail.com> Co-authored-by: Xiaoyu Zhang <1182563586@qq.com>
This commit is contained in:
co-authored by
guzekai01
Julian Huang
墨楼
Claude Opus 4.8
Peng Zhang
Xiaoyu Zhang
parent
4c997310f5
commit
423b8485fb
@@ -779,6 +779,26 @@ SGLang supports various environment variables that can be used to configure its
|
||||
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>A comma-separated list of layer names to keep out of FP4 online quantization, including <code>nvfp4_online</code>. For example: <code>model.layers.40,model.layers.41</code>.</td>
|
||||
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}><code>""</code></td>
|
||||
</tr>
|
||||
<tr>
|
||||
<td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>SGLANG_HUMMING_ONLINE_QUANT_CONFIG</code></td>
|
||||
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>JSON object or JSON file path for Humming online weight quantization. When a layer has no checkpoint quantization config, this config tells Humming how to quantize the loaded fp16/bf16 weight. When the checkpoint already has a Humming config, add <code>"force_requant": true</code> to requantize it to this schema during loading. Common keys include <code>dtype</code>/<code>weight_dtype</code>, <code>scale_dtype</code>, <code>group_size</code>, <code>scale_type</code>, <code>ignored_layers</code>, <code>ignore</code>, and <code>modules_to_not_convert</code>.</td>
|
||||
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}><code>None</code></td>
|
||||
</tr>
|
||||
<tr>
|
||||
<td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>SGLANG_HUMMING_INPUT_QUANT_CONFIG</code></td>
|
||||
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>JSON object or JSON file path for Humming input activation quantization. This controls the activation dtype and scale grouping passed into Humming kernels, independently of the weight schema. For example, <code>{'{"dtype": "float8e4m3"}'}</code> quantizes Humming inputs to FP8 E4M3.</td>
|
||||
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}><code>None</code></td>
|
||||
</tr>
|
||||
<tr>
|
||||
<td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>SGLANG_HUMMING_USE_F16_ACCUM</code></td>
|
||||
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Use FP16 accumulation in Humming compute/tuning config. This is only meaningful for Humming dtype combinations that support FP16 accumulation, such as fp16 or FP8 E4M3 activations with float16 output. Leave it <code>false</code> for the default accumulator behavior.</td>
|
||||
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}><code>false</code></td>
|
||||
</tr>
|
||||
<tr>
|
||||
<td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>SGLANG_HUMMING_MOE_GEMM_TYPE</code></td>
|
||||
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Select the Humming MoE GEMM path for standard dispatch and DeepEP normal dispatch. <code>indexed</code> uses top-k expert ids directly and is the fallback for unset or unknown values. <code>grouped</code> maps to Humming grouped-contiguous GEMM. DeepEP low-latency dispatch uses grouped-masked GEMM internally and does not use this selector.</td>
|
||||
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}><code>""</code> (<code>indexed</code>)</td>
|
||||
</tr>
|
||||
</tbody>
|
||||
</table>
|
||||
|
||||
|
||||
Reference in New Issue
Block a user