[Bugfix][Spec Decode] Qualify Transformers backend attention layer names with the model prefix (#59293)

Signed-off-by: Farhad Al-Amin <99006654+Legend-2727@users.noreply.github.com>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Farhad Al-Amin
2026-10-02 11:21:20 +01:00
committed by GitHub
co-authored by Claude Opus 5.5 Harry Mellor
parent 8e58ac22af
commit af2a582d5c
2 changed files with 24 additions and 8 deletions
@@ -150,6 +150,21 @@ def test_draft_model_quantization(
assert_draft_model_correctness(sd_case, vllm_runner)
@single_gpu_only
def test_draft_model_transformers_backend(vllm_runner):
"""Ensure that Attention prefixes don't collide between the target and
draft models when both models use the Transformers modelling backend."""
sd_case = ArgsTest(
target_model="allenai/OLMo-2-0425-1B-Instruct",
draft_model="allenai/OLMo-2-0425-1B-Instruct",
sampling_config=greedy_sampling(),
num_speculative_tokens=3,
expected_acceptance_len=0.98 * (3 + 1), # epsilon discount of K + 1
expected_acceptance_rate=0.98, # slight epsilon
)
assert_draft_model_correctness(sd_case, vllm_runner)
@multi_gpu_only(num_gpus=2)
def test_draft_model_tensor_parallelism(vllm_runner):
"""Ensure spec decode works when running with TP > 1."""