[ROCm] Remove stale SDPA and skinny GEMM workarounds (#50907)

Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Andreas Karatzas
2026-08-11 14:19:16 -04:00
committed by GitHub
co-authored by OpenAI Codex
parent 4f2f31b82e
commit 1ab2801dde
56 changed files with 500 additions and 558 deletions
+20
View File
@@ -50,6 +50,20 @@ def reformat(text: str) -> str:
return text
def _enable_deterministic_lora_shrink(monkeypatch: pytest.MonkeyPatch) -> None:
if not current_platform.is_rocm():
return
# These tests assert exact greedy outputs. Force the Triton LoRA shrink
# kernel to use SPLIT_K=1 so it stores the complete reduction directly
# instead of accumulating split-K partial results with atomic_add.
monkeypatch.setenv("VLLM_BATCH_INVARIANT", "1")
# The kernel configuration reads VLLM_BATCH_INVARIANT at import time.
# Spawn the engine process so it observes this setting even if the LoRA
# Triton utilities were already imported during test collection.
monkeypatch.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
def generate_and_test(llm: vllm.LLM, lora_path: str, lora_id: int) -> None:
prompts = [
PROMPT_TEMPLATE.format(
@@ -106,7 +120,10 @@ def test_gpt_oss_lora(
gptoss20b_lora_files,
mxfp4_use_marlin,
specialize_active_lora,
monkeypatch: pytest.MonkeyPatch,
):
_enable_deterministic_lora_shrink(monkeypatch)
llm = vllm.LLM(
MODEL_PATH,
max_model_len=1024,
@@ -146,7 +163,10 @@ def test_gpt_oss_lora_tp2(
gptoss20b_lora_files,
fully_sharded_loras,
mxfp4_use_marlin,
monkeypatch: pytest.MonkeyPatch,
):
_enable_deterministic_lora_shrink(monkeypatch)
llm = vllm.LLM(
MODEL_PATH,
max_model_len=1024,
+21
View File
@@ -40,6 +40,18 @@ def sample_data(num_experts, max_loras, num_tokens, topk_num):
@pytest.mark.parametrize("block_size", [16])
def test_moe_lora_align_block_size(
num_tokens, topk_num, num_experts, max_loras, block_size
):
_run_moe_lora_align_block_size_case(
num_tokens, topk_num, num_experts, max_loras, block_size
)
def test_moe_lora_align_block_size_large_path_counts():
_run_moe_lora_align_block_size_case(100, 12, 64, 4, 16)
def _run_moe_lora_align_block_size_case(
num_tokens, topk_num, num_experts, max_loras, block_size
):
# sample data
random.seed(1)
@@ -94,8 +106,17 @@ def test_moe_lora_align_block_size(
# verify values
expert_ids = expert_ids.view(max_loras, -1)
sorted_token_ids = sorted_token_ids.view(max_loras, -1, block_size)
topk_ids_cpu = topk_ids.cpu()
token_lora_mapping_cpu = token_lora_mapping.cpu()
for lora_idx in range(max_loras):
expert_counts = torch.bincount(
topk_ids_cpu[token_lora_mapping_cpu == lora_idx].flatten(),
minlength=num_experts,
)
expected_post_pad = round_up(expert_counts, block_size).sum().item()
assert num_tokens_post_pad[lora_idx].item() == expected_post_pad
for token_idx in range(sorted_token_ids.size(1)):
block = sorted_token_ids[lora_idx][token_idx]
indices = block[block != topk_ids.numel()]