[Core][LoRA] Support fp32 lm_head (head_dtype) on the LoRA path (#48525)

Signed-off-by: Karthik Kothuri <karthikkothuri2009@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
Karthik Kothuri
2026-07-15 09:51:09 +08:00
committed by GitHub
co-authored by Claude
parent 6e073440b1
commit 4f7fffb92f
3 changed files with 11 additions and 29 deletions
+4 -1
View File
@@ -368,8 +368,9 @@ def test_embeddings(
@pytest.mark.parametrize("device", DEVICES)
@pytest.mark.parametrize("vocab_size", [64000, 256512, 258048])
@pytest.mark.parametrize("stage", STAGES)
@pytest.mark.parametrize("head_dtype", [None, torch.float32])
def test_lm_head_logits_processor(
default_vllm_config, dist_init, num_loras, device, vocab_size, stage
default_vllm_config, dist_init, num_loras, device, vocab_size, stage, head_dtype
) -> None:
if current_platform.is_cuda_alike() or current_platform.is_xpu():
torch.accelerator.set_device_index(device)
@@ -391,6 +392,8 @@ def test_lm_head_logits_processor(
linear.weight.data = torch.rand_like(linear.weight.data)
linear.weight.data[:, vocab_size:] = 0
logits_processor = LogitsProcessor(vocab_size)
# Exercise an fp32 lm_head (head_dtype) on the LoRA path.
logits_processor.head_dtype = head_dtype
lora_logits_processor = LogitsProcessorWithLoRA(
logits_processor, 1024, linear.weight.dtype, linear.weight.device, None
)
-16
View File
@@ -136,22 +136,6 @@ def test_get_top_tokens_honors_head_dtype(default_vllm_config):
assert torch.equal(top, expected)
def test_fp32_head_rejected_with_lora(default_vllm_config):
from vllm.lora.layers.logits_processor import LogitsProcessorWithLoRA
base = _build_processor(64)
base.head_dtype = torch.float32
with pytest.raises(ValueError, match="not yet supported with LoRA"):
LogitsProcessorWithLoRA(
base,
hidden_size=16,
dtype=torch.bfloat16,
device=torch.device("cpu"),
sharded_to_full_mapping=None,
)
@pytest.mark.core_model
def test_fp32_head_e2e_no_nan():
"""An fp32 head produces finite logprobs end-to-end.