diff --git a/tests/kernels/attention/test_minimax_m3.py b/tests/kernels/attention/test_minimax_m3.py index 7ee900fd3d..918e4562ce 100644 --- a/tests/kernels/attention/test_minimax_m3.py +++ b/tests/kernels/attention/test_minimax_m3.py @@ -1833,6 +1833,35 @@ def test_aiter_indexer_requires_the_aiter_attend(monkeypatch): assert reason is None or "AITER sparse PA attend" not in reason +def test_aiter_indexer_rejects_unbuilt_index_heads(monkeypatch): + """AITER builds the MSA score kernels for 1 or 2 index heads only. + + MiniMax-M3 has 4, so at attention TP=1 the indexer has to fall back at load + time rather than fail at the first forward. + """ + import vllm.models.minimax_m3.amd.indexer_aiter as indexer_aiter_mod + + monkeypatch.setattr(indexer_aiter_mod.current_platform, "is_rocm", lambda: True) + monkeypatch.setattr( + indexer_aiter_mod, "_minimax_m3_aiter_sparse_pa_requested", lambda: True + ) + + def unsupported_reason(num_index_heads: int) -> str | None: + return indexer_aiter_mod.aiter_indexer_unsupported_reason( + topk_blocks=TOPK, + sparse_block_size=BLOCK_SIZE, + num_index_heads=num_index_heads, + index_head_dim=HEAD_DIM, + indexer_kv_dtype="fp8", + max_model_len=8192, + ) + + if reason := unsupported_reason(2): + pytest.skip(f"AITER MSA indexer unavailable: {reason}") + reason = unsupported_reason(4) + assert reason is not None and "num_index_heads" in reason + + def test_aiter_consolidated_qknorm_enabled(monkeypatch): from vllm._aiter_ops import rocm_aiter_ops diff --git a/vllm/models/minimax_m3/amd/indexer_aiter.py b/vllm/models/minimax_m3/amd/indexer_aiter.py index d46db51363..f0184a5c6a 100644 --- a/vllm/models/minimax_m3/amd/indexer_aiter.py +++ b/vllm/models/minimax_m3/amd/indexer_aiter.py @@ -67,6 +67,8 @@ MSA_TOPK_BLOCKS = 16 MSA_SCORE_TYPE = "max" MSA_SPARSE_BLOCK_SIZE = 128 MSA_INDEX_HEAD_DIM = 128 +# Index head counts the score kernels are instantiated for. +MSA_NUM_INDEX_HEADS = (1, 2) # Wave width the top-k is written against: it gives one lane per output slot and # reads the score row in wave-wide strips. WAVE_SIZE = 64 @@ -181,8 +183,11 @@ def aiter_indexer_unsupported_reason( f"needs index_head_dim={MSA_INDEX_HEAD_DIM}, " f"got index_head_dim={index_head_dim}" ) - if num_index_heads > MFMA_COLS: - return f"num_index_heads={num_index_heads} exceeds the {MFMA_COLS} MFMA columns" + if num_index_heads not in MSA_NUM_INDEX_HEADS: + return ( + f"needs num_index_heads in {MSA_NUM_INDEX_HEADS}, " + f"got num_index_heads={num_index_heads}" + ) # A decode row's whole query shares one MFMA tile, one column per (token, # head) pair, so spec decode trades columns against the head count. if num_index_heads * max_decode_query_len > MFMA_COLS: diff --git a/vllm/models/minimax_m3/common/sparse_attention.py b/vllm/models/minimax_m3/common/sparse_attention.py index 61791dc5ff..1346765f84 100644 --- a/vllm/models/minimax_m3/common/sparse_attention.py +++ b/vllm/models/minimax_m3/common/sparse_attention.py @@ -125,7 +125,9 @@ def minimax_m3_use_aiter_sparse_pa( raise ValueError( "MiniMax M3 AITER sparse paged attention requires " f"num_kv_heads == 1 per tensor-parallel rank, got {num_kv_heads}, " - "unless the indexer emits the attend's page table." + "unless the AITER indexer emits the attend's page table. It is not " + "in use here; the 'MiniMax M3 indexer: AITER unavailable' log line " + "says why." ) return requested