A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 04:45 UTC
[CI Fix] Try fixing eagle e2e test OOM by reducing block allocation (#20213)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Jun 29, 2025 at 11:31 UTC
7b1895e6ce4942091e16da790af8c12772a1d384
1 file changed
+8
tests/spec_decode/e2e/test_eagle_correctness.py
+8
@@ -370,6 +370,10 @@ def test_llama2_eagle_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
370
@pytest.mark.parametrize(
371
"common_llm_kwargs",
372
[{
373
+ # 2 for small prompt, 256//16 for generated.
374
+ "num_gpu_blocks_override": 2 + 256 // 16,
375
+ "max_model_len": (2 + 256 // 16) * 16,
376
+
377
# Skip cuda graph recording for fast test.
378
"enforce_eager": True,
379
@@ -420,6 +424,10 @@ def test_llama3_eagle_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
424
@pytest.mark.parametrize(
425
"common_llm_kwargs",
426
[{
427
+ # 2 for small prompt, 256//16 for generated.
428
+ "num_gpu_blocks_override": 2 + 256 // 16,
429
+ "max_model_len": (2 + 256 // 16) * 16,
430
+
431
# Skip cuda graph recording for fast test.
432
"enforce_eager": True,
433