minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Qwen3.8] Enable NVIDIA NVFP4 on DGX Spark with file-backed PLE and PDL router fix (#39126)

Browse Source 
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: rdxa <rdxa@rdxa-int-spark-01.yvb.moe>
Co-authored-by: Yangmin Li <yangminl@nvidia.com>
Co-authored-by: Manrique <nanomlm@gmail.com>
Co-authored-by: yhyang201 <yhyang201@gmail.com>
This commit is contained in:
Jimmy Shong
2026-09-13 16:23:41 +08:00
committed by GitHub
co-authored by Claude Fable 5.1 rdxa Yangmin Li Manrique yhyang201
parent d6fabb74b4
commit cebca698e2
21 changed files with 1376 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/unit/model_loader/test_modelopt_loader.py
+56 -1
View File
@@ -22,7 +22,12 @@ from sglang.srt.configs.model_config import ModelConfig
from sglang.srt.layers.linear import ReplicatedLinear
from sglang.srt.layers.logits_processor import should_apply_lm_head_quant_method
from sglang.srt.layers.modelopt_utils import QUANT_CFG_CHOICES
from sglang.srt.layers.quantization.fp8 import Fp8Config, Fp8LinearMethod
from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
from sglang.srt.layers.quantization.fp8 import (
Fp8Config,
Fp8LinearMethod,
Fp8MoEMethod,
)
from sglang.srt.layers.quantization.modelopt_quant import (
ModelOptFp4Config,
ModelOptFp4LinearMethod,
@@ -1188,6 +1193,56 @@ class TestModelOptMixedPrecisionConfig(CustomTestCase):
"FP8",
)
def test_mixed_precision_resolves_vl_language_model_keys(self):
# nvidia/Qwen3.8-Flash-Next-NVFP4 keys the text stack as
# `model.language_model.*` while Qwen4-Exp modules are `model.*`.
quant_config = ModelOptMixedPrecisionConfig.from_config(
{
"quant_algo": "MIXED_PRECISION",
"quantized_layers": {
"model.language_model.layers.3.mlp.experts": {
"quant_algo": "NVFP4",
"group_size": 16,
},
"model.language_model.layers.1.ple.ple_embedding.ngram_embedding": {
"quant_algo": "FP8"
},
"mtp.layers.0.mlp.experts": {
"quant_algo": "FP8_BLOCK_SCALES",
"group_size": 128,
},
},
}
)
self.assertEqual(quant_config.exclude_modules, [])
moe = FusedMoE.__new__(FusedMoE)
self.assertIsInstance(
quant_config.get_quant_method(moe, "mtp.layers.0.mlp.experts"),
Fp8MoEMethod,
)
self.assertEqual(
quant_config.get_quant_method(
moe, "mtp.layers.0.mlp.experts"
).quant_config.weight_block_size,
[128, 128],
)
self.assertEqual(
quant_config.resolve_quant_algo("model.layers.3.mlp.experts"), "NVFP4"
)
self.assertEqual(
quant_config.resolve_quant_algo(
"model.layers.1.ple.ple_embedding.ngram_embedding"
),
"FP8",
)
self.assertIsNone(
quant_config.resolve_quant_algo("model.layers.1.ple.key_proj")
)
self.assertIsNone(
quant_config.resolve_quant_algo("model.layers.3.mlp.shared_expert")
)
if __name__ == "__main__":
unittest.main()