test cleanup

Signed-off-by: Bill Nell <bnell@redhat.com>

Author: bnellnm

tests/kernels/moe/test_block_fp8.py

@@ -45,7 +45,7 @@
 GROUP_SIZE = [64, 128, 512]
 # Deepseek-V3's intermediate size 18432, so N is 18432*2/8=4608 at TP8
 # and its hidden size is 7168.
-M = [1, 2, 83, 128, 2048, 1024 * 128]
+M = [1, 2, 83, 128, 2048, 32768]
 M_dg = [128, 192, 1335, 2048]
 N = [128, 256, 1024, 4608]  # [13824]
 K = [256, 512, 7168]  # [13824]
@@ -104,11 +104,15 @@ def setup_cuda():
     "M,N,K,E,topk,block_size,dtype,seed",
     itertools.product(M, N, K, E, TOP_KS, BLOCK_SIZE, DTYPES, SEEDS))
 @torch.inference_mode()
-def test_w8a8_block_fp8_fused_moe(M, N, K, E, topk, block_size, dtype, seed):
+def test_w8a8_block_fp8_fused_moe(M, N, K, E, topk, block_size, dtype, seed,
+                                  monkeypatch):
     if topk > E:
         pytest.skip(f"Skipping test; topk={topk} > E={E}")
 
     torch.manual_seed(seed)
+
+    monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
+
     factor_for_scale = 1e-2
     fp8_info = torch.finfo(torch.float8_e4m3fn)
     fp8_max, fp8_min = fp8_info.max, fp8_info.min
@@ -262,19 +266,20 @@ def deep_gemm_w8a8_block_fp8_moe(M, K, a, w1, w2, w1_s, w2_s, score, topk,
     itertools.product(M_dg, N, K, E, TOP_KS, SEEDS))
 @pytest.mark.skipif(not dg_available, reason="DeepGemm kernels not available.")
 @torch.inference_mode()
-def test_w8a8_block_fp8_deep_gemm_fused_moe(M, N, K, E, topk, seed):
-
-    block_m = deep_gemm.get_m_alignment_for_contiguous_layout()
-    block_size = [block_m, block_m]
-    dtype = torch.bfloat16
-
+def test_w8a8_block_fp8_deep_gemm_fused_moe(M, N, K, E, topk, seed,
+                                            monkeypatch):
     if topk > E:
         pytest.skip(f"Skipping test: topk={topk} > E={E}")
 
     if not _valid_deep_gemm_shape(M, N, K):
         pytest.skip(f"Skipping test: invalid size m={M}, n={N}, k={K}")
 
     torch.manual_seed(seed)
+
+    monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
+    block_m = deep_gemm.get_m_alignment_for_contiguous_layout()
+    block_size = [block_m, block_m]
+    dtype = torch.bfloat16
     fp8_info = torch.finfo(torch.float8_e4m3fn)
     fp8_max, fp8_min = fp8_info.max, fp8_info.min
 

tests/kernels/moe/test_cutlass_moe.py

@@ -29,7 +29,7 @@
     (224, 1024, 1536),
     (224, 3072, 1024),
     (224, 3072, 1536),
-    (1024 * 128, 1024, 1024),
+    (32768, 1024, 1024),
 ]
 
 vllm_config = VllmConfig(parallel_config=ParallelConfig(
@@ -232,8 +232,10 @@ def test_cutlass_moe_8_bit_no_graph(
     topk: int,
     per_act_token: bool,
     per_out_ch: bool,
+    monkeypatch,
 ):
     current_platform.seed_everything(7)
+    monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
     with set_current_vllm_config(vllm_config):
         mt = MOETensors8Bit.make_moe_tensors_8bit(m, k, n, e, per_act_token,
                                                   per_out_ch)
@@ -274,8 +276,10 @@ def test_cutlass_moe_8_bit_cuda_graph(
     topk: int,
     per_act_token: bool,
     per_out_ch: bool,
+    monkeypatch,
 ):
     current_platform.seed_everything(7)
+    monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
     with set_current_vllm_config(vllm_config):
         dtype = torch.half
 
@@ -329,8 +333,10 @@ def test_cutlass_moe_8_bit_EP(
     per_act_token: bool,
     per_out_channel: bool,
     ep_size: int,
+    monkeypatch,
 ):
     current_platform.seed_everything(7)
+    monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
     with set_current_vllm_config(vllm_config):
         mt = MOETensors8Bit.make_moe_tensors_8bit(m, k, n, e, per_act_token,
                                                   per_out_channel)

tests/kernels/moe/test_moe.py

@@ -40,7 +40,7 @@
 vllm_config.scheduler_config.max_model_len = 8192
 
 
-@pytest.mark.parametrize("m", [1, 33, 64, 222, 1024 * 128])
+@pytest.mark.parametrize("m", [1, 33, 64, 222, 32768])
 @pytest.mark.parametrize("n", [128, 1024, 2048])
 @pytest.mark.parametrize("k", [128, 511, 1024])
 @pytest.mark.parametrize("e", NUM_EXPERTS)
@@ -57,7 +57,11 @@ def test_fused_moe(
     ep_size: int,
     dtype: torch.dtype,
     padding: bool,
+    monkeypatch,
 ):
+    current_platform.seed_everything(7)
+    monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
+
     a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
     w1 = torch.randn((e, 2 * n, k), device="cuda", dtype=dtype) / 10
     w2 = torch.randn((e, k, n), device="cuda", dtype=dtype) / 10
@@ -81,7 +85,7 @@ def test_fused_moe(
                                            use_int8_w8a8=False,
                                            use_int8_w8a16=False,
                                            use_int4_w4a16=False,
-                                           per_channel_quant=False,
+                                           per_act_token_quant=False,
                                            block_shape=None)
 
     with set_current_vllm_config(vllm_config):