[ROCm][CI] Fix test_max_len.py for Rocm

[charlifu]

Use ROCM_AITER_FA backend for flash attn on rocm.

[gemini-code-assist]

Code Review

This pull request updates the attention backend for ROCm in the test utilities, changing it from FLASH_ATTN to ROCM_AITER_FA. This aligns with using the Aiter Flash Attention backend on ROCm platforms. My review includes a suggestion to update a related log message for consistency and to prevent confusion during debugging.

[divakar-amd]

Hi @charlifu , can you also update it here:

vllm/tests/v1/e2e/test_spec_decode.py

Lines 400 to 418 in 1c593e1

	with monkeypatch.context() as m:
	if "Llama-4-Scout" in model_setup[1] and attn_backend == "FLASH_ATTN":
	# Scout requires default backend selection
	# because vision encoder has head_dim 88 being incompatible
	# with FLASH_ATTN and needs to fall back to Flex Attn
	pass
	else:
	m.setenv("VLLM_MLA_DISABLE", "1")
	m.setenv("VLLM_ATTENTION_BACKEND", attn_backend)
	if attn_backend == "TRITON_ATTN" and not current_platform.is_rocm():
	pytest.skip(
	"TRITON_ATTN does not support "
	"multi-token eagle spec decode on current platform"
	)
	if attn_backend == "FLASH_ATTN" and current_platform.is_rocm():
	if "deepseek" in model_setup[1].lower():
	pytest.skip("FLASH_ATTN for deepseek not supported on ROCm platform")

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @charlifu.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[tjtanaa]

@charlifu after you done fixing it, can you share the tests status of the files that you have fixed?

[tjtanaa]

on mi300x, I am getting OOM issues on two of the tests. May I know which machine are you validating on? mi325?

[robertgshaw2-redhat]

removing approval since @tjtanaa is on point

[charlifu]

@charlifu after you done fixing it, can you share the tests status of the files that you have fixed?

pytest -sv tests/v1/spec_decode/test_max_len.py

9 passed, 4 warnings in 202.83s (0:03:22)

pytest -sv tests/v1/spec_decode/test_eagle.py

51 passed, 4 warnings in 37.66s

[tjtanaa]

@charlifu how about the status of tests/v1/e2e/test_spec_decode.py ?

[charlifu]

@charlifu how about the status of tests/v1/e2e/test_spec_decode.py ?

Still failing, but that's not what this PR aiming for. Can we get this merged first?

[micah-wil]

@charlifu how about the status of tests/v1/e2e/test_spec_decode.py ?

Fails for unrelated reason that is addressed in #29926

[charlifu]

Another fix here:

pytest -sv tests/basic_correctness/test_basic_correctness.py

=============== 34 passed, 6 warnings in 480.39s (0:08:00) ==========

[tjtanaa]

LGTM. Let's not any more tests. Open another PR if there are new tests.

[tjtanaa]

@micah-wil I saw you have merged the bugfix PR into this PR by syncing from main. Are the test cases in tests/v1/e2e/test_spec_decode.py all green? I think this test is yet to be enabled in AMD CI. Can you share the status of tests/v1/e2e/test_spec_decode.py.