[Hardware] Replace torch.cuda.synchronize() api with torch.accelerator.synchronize (#36085)

Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
Kunshang Ji
2026-03-05 10:36:39 +00:00
committed by GitHub
parent 0bfa229bf1
commit 66a2209645
59 changed files with 158 additions and 161 deletions
@@ -34,7 +34,7 @@ def do_profile(
record_shapes=True,
) as tprof:
fn(**fn_kwargs)
torch.cuda.synchronize(torch.cuda.current_device())
torch.accelerator.synchronize(torch.cuda.current_device())
# TODO (varun): Add a descriptive trace file name
tprof.export_chrome_trace(
+2 -2
View File
@@ -318,8 +318,8 @@ def test_w8a8_block_fp8_deep_gemm_fused_moe(M, N, K, E, topk, seed, monkeypatch)
out = deep_gemm_moe_fp8_fn(
a, w1, w2, w1_s, w2_s, topk_weights, topk_ids
)
torch.cuda.synchronize()
torch.accelerator.synchronize()
graph.replay()
torch.cuda.synchronize()
torch.accelerator.synchronize()
torch.testing.assert_close(out, ref_out, atol=0.035, rtol=0.035)
+2 -2
View File
@@ -399,9 +399,9 @@ def test_cutlass_moe_8_bit_cuda_graph(
mt, topk_weights, topk_ids, per_act_token, per_out_ch
)
torch.cuda.synchronize()
torch.accelerator.synchronize()
graph.replay()
torch.cuda.synchronize()
torch.accelerator.synchronize()
torch.testing.assert_close(triton_output, cutlass_output, atol=9e-2, rtol=1e-2)
+3 -3
View File
@@ -272,9 +272,9 @@ def run_moe_test(
global_num_experts=global_num_experts,
expert_map=expert_map,
)
torch.cuda.synchronize()
torch.accelerator.synchronize()
graph.replay()
torch.cuda.synchronize()
torch.accelerator.synchronize()
torch.testing.assert_close(test_output, baseline_output, atol=atol, rtol=rtol)
@@ -768,7 +768,7 @@ def test_mixtral_moe(
F.pad(vllm_moe.experts.w2_weight, (0, 128), "constant", 0)[..., 0:-128],
requires_grad=False,
)
torch.cuda.synchronize()
torch.accelerator.synchronize()
torch.accelerator.empty_cache()
# FIXME (zyongye) fix this after we move self.kernel