Add LM head quantization support for ModelOpt (#42124)

Signed-off-by: weimingc <17592131+meenchen@users.noreply.github.com>
This commit is contained in:
Wei-Ming Chen
2026-05-26 09:21:05 -07:00
committed by GitHub
parent c8414a8271
commit 6f5b533241
8 changed files with 220 additions and 5 deletions
@@ -0,0 +1,34 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from unittest.mock import Mock, patch
def test_nemotron_h_lm_head_receives_quant_config():
from vllm.model_executor.models.nemotron_h import NemotronHForCausalLM
mock_quant_config = Mock()
mock_hf_config = Mock()
mock_hf_config.vocab_size = 128
mock_hf_config.hidden_size = 64
mock_vllm_config = Mock()
mock_vllm_config.model_config.hf_config = mock_hf_config
mock_vllm_config.model_config.dtype = None
mock_vllm_config.scheduler_config = Mock()
mock_vllm_config.quant_config = mock_quant_config
with (
patch("vllm.model_executor.models.nemotron_h.NemotronHModel") as MockModel,
patch("vllm.model_executor.models.nemotron_h.ParallelLMHead") as MockLMHead,
patch("vllm.model_executor.models.nemotron_h.LogitsProcessor"),
):
MockModel.return_value.make_empty_intermediate_tensors = Mock()
MockModel.return_value.has_moe = False
NemotronHForCausalLM(vllm_config=mock_vllm_config)
MockLMHead.assert_called_once()
call_kwargs = MockLMHead.call_args.kwargs
assert call_kwargs["quant_config"] is mock_quant_config
@@ -0,0 +1,78 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from unittest.mock import Mock, patch
def test_qwen3_5_lm_head_receives_quant_config():
from vllm.model_executor.models.qwen3_5 import Qwen3_5ForCausalLMBase
mock_quant_config = Mock()
mock_hf_config = Mock()
mock_hf_config.tie_word_embeddings = False
mock_hf_config.vocab_size = 128
mock_hf_config.hidden_size = 64
mock_vllm_config = Mock()
mock_vllm_config.model_config.hf_text_config = mock_hf_config
mock_vllm_config.cache_config.mamba_cache_mode = "align"
mock_vllm_config.scheduler_config = Mock()
mock_vllm_config.quant_config = mock_quant_config
mock_vllm_config.lora_config = None
mock_pp_group = Mock()
mock_pp_group.is_last_rank = True
with (
patch("vllm.model_executor.models.qwen3_5.Qwen3_5Model") as MockModel,
patch("vllm.model_executor.models.qwen3_5.ParallelLMHead") as MockLMHead,
patch("vllm.model_executor.models.qwen3_5.LogitsProcessor"),
patch(
"vllm.model_executor.models.qwen3_5.get_pp_group",
return_value=mock_pp_group,
),
):
MockModel.return_value.make_empty_intermediate_tensors = Mock()
Qwen3_5ForCausalLMBase(vllm_config=mock_vllm_config)
MockLMHead.assert_called_once()
call_kwargs = MockLMHead.call_args.kwargs
assert call_kwargs["quant_config"] is mock_quant_config
def test_qwen3_5_mtp_lm_head_receives_quant_config():
from vllm.config import CompilationMode
from vllm.model_executor.models.qwen3_5_mtp import Qwen3_5MTP
mock_quant_config = Mock()
mock_hf_config = Mock()
mock_hf_config.tie_word_embeddings = False
mock_hf_config.vocab_size = 128
mock_hf_config.hidden_size = 64
mock_vllm_config = Mock()
mock_vllm_config.model_config.hf_text_config = mock_hf_config
mock_vllm_config.cache_config.mamba_cache_mode = "align"
mock_vllm_config.compilation_config.mode = CompilationMode.NONE
mock_vllm_config.quant_config = mock_quant_config
mock_pp_group = Mock()
mock_pp_group.is_last_rank = True
with (
patch("vllm.model_executor.models.qwen3_5_mtp.Qwen3_5MultiTokenPredictor"),
patch("vllm.model_executor.models.qwen3_5_mtp.ParallelLMHead") as MockLMHead,
patch("vllm.model_executor.models.qwen3_5_mtp.LogitsProcessor"),
patch(
"vllm.model_executor.models.qwen3_5_mtp.get_pp_group",
return_value=mock_pp_group,
),
):
Qwen3_5MTP(vllm_config=mock_vllm_config)
MockLMHead.assert_called_once()
call_kwargs = MockLMHead.call_args.kwargs
assert call_kwargs["quant_config"] is mock_quant_config