Support quantized DSpark Markov heads (#50424)
Signed-off-by: Andrii Skliar <askliar@nvidia.com> Co-authored-by: Andrii Skliar <askliar@nvidia.com> Co-authored-by: Benjamin Chislett <bchislett@nvidia.com>
Andrii Skliar committed
Aug 3, 2026 at 14:35 UTC
b977407d8be3bc076c1e2d6f3df02b9bdbd1fab9
1 file changed
+4
vllm/model_executor/models/qwen3_dspark.py
+4
@@ -22,6 +22,7 @@ import torch.nn as nn
22
from vllm.config import VllmConfig
23
from vllm.logger import init_logger
24
from vllm.model_executor.layers.logits_processor import LogitsProcessor
25
+from vllm.model_executor.layers.quantization import QuantizationConfig
26
from vllm.model_executor.layers.vocab_parallel_embedding import (
27
ParallelLMHead,
28
)
@@ -51,6 +52,7 @@ class DSparkMarkovHead(nn.Module):
52
draft_vocab_size: int,
53
markov_rank: int,
54
prefix: str,
55
+ quant_config: QuantizationConfig | None = None,
56
) -> None:
57
super().__init__()
58
self.markov_w1 = nn.Embedding(vocab_size, markov_rank)
@@ -58,6 +60,7 @@ class DSparkMarkovHead(nn.Module):
60
draft_vocab_size,
61
markov_rank,
62
bias=False,
63
+ quant_config=quant_config,
64
prefix=maybe_prefix(prefix, "markov_w2"),
65
disable_tp=True,
66
)
@@ -97,6 +100,7 @@ class Qwen3DSparkModel(DFlashQwen3Model):
100
draft_vocab_size,
101
config.markov_rank,
102
prefix=maybe_prefix(prefix, "markov_head"),
103
+ quant_config=self.quant_config,
104
)
105
106