the draft-routing patch enabling DSpark on NVFP4 checkpoints
Browse files
dspark-nvfp4-draft-routing.patch
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
diff --git a/vllm/models/deepseek_v4/quant_config.py b/vllm/models/deepseek_v4/quant_config.py
|
| 2 |
+
index 89cf695..5918fa0 100644
|
| 3 |
+
--- a/vllm/models/deepseek_v4/quant_config.py
|
| 4 |
+
+++ b/vllm/models/deepseek_v4/quant_config.py
|
| 5 |
+
@@ -4,6 +4,7 @@
|
| 6 |
+
|
| 7 |
+
from __future__ import annotations
|
| 8 |
+
|
| 9 |
+
+import re
|
| 10 |
+
from typing import TYPE_CHECKING, cast
|
| 11 |
+
|
| 12 |
+
from vllm.config import get_current_vllm_config
|
| 13 |
+
@@ -175,7 +176,7 @@ class DeepseekV4FP8Config(Fp8Config):
|
| 14 |
+
):
|
| 15 |
+
return UnquantizedFusedMoEMethod(layer.moe_config)
|
| 16 |
+
if self.expert_dtype == "fp4":
|
| 17 |
+
- if self.moe_quant_algo == "NVFP4":
|
| 18 |
+
+ if self.moe_quant_algo == "NVFP4" and not self._is_draft_layer(prefix):
|
| 19 |
+
from vllm.model_executor.layers.quantization.modelopt import (
|
| 20 |
+
ModelOptNvFp4FusedMoE,
|
| 21 |
+
)
|
| 22 |
+
@@ -189,7 +190,20 @@ class DeepseekV4FP8Config(Fp8Config):
|
| 23 |
+
# returns Fp8MoEMethod with block-wise float32 scales.
|
| 24 |
+
return super().get_quant_method(layer, prefix)
|
| 25 |
+
|
| 26 |
+
+ def _is_draft_layer(self, prefix: str) -> bool:
|
| 27 |
+
+ # NVFP4 conversions quantize only the main stack; the speculative
|
| 28 |
+
+ # module (runtime layer index >= num_hidden_layers) keeps its
|
| 29 |
+
+ # original MXFP4 experts and must load through the MXFP4 path.
|
| 30 |
+
+ m = re.search(r"layers\.(\d+)\.", prefix)
|
| 31 |
+
+ if m is None:
|
| 32 |
+
+ return False
|
| 33 |
+
+ try:
|
| 34 |
+
+ hf_config = get_current_vllm_config().model_config.hf_config
|
| 35 |
+
+ except Exception:
|
| 36 |
+
+ return False
|
| 37 |
+
+ return int(m.group(1)) >= hf_config.num_hidden_layers
|
| 38 |
+
+
|
| 39 |
+
def is_mxfp4_quant(self, prefix, layer):
|
| 40 |
+
if not isinstance(layer, RoutedExperts) or self.expert_dtype != "fp4":
|
| 41 |
+
return False
|
| 42 |
+
- return self.moe_quant_algo != "NVFP4"
|
| 43 |
+
+ return self.moe_quant_algo != "NVFP4" or self._is_draft_layer(prefix)
|