chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[CI] Move existing unit tests into unit directory (#20631)

Browse Source
This commit is contained in:
Ke Bao
2026-03-15 23:25:18 +08:00
committed by GitHub
parent e2be31824f
commit c3483e8e97
29 changed files with 16 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/unit/layers/test_conv_layer.py
+367
View File
@@ -0,0 +1,367 @@
from sglang.test.ci.ci_register import register_cuda_ci
register_cuda_ci(est_time=10, suite="stage-b-test-small-1-gpu")
import unittest
import torch
import torch.nn as nn
from sglang.srt.layers.conv import Conv2dLayer, Conv3dLayer
def _copy_weights(src, dst_nn):
"""Copy weights from Conv*dLayer to nn.Conv*d for comparison."""
with torch.no_grad():
dst_nn.weight.copy_(src.weight)
if src.bias is not None:
dst_nn.bias.copy_(src.bias)
class TestConv2dLayer(unittest.TestCase):
def test_basic_patch_embedding(self):
layer = Conv2dLayer(3, 768, kernel_size=14, stride=14, bias=False)
ref = nn.Conv2d(3, 768, kernel_size=14, stride=14, bias=False)
self.assertFalse(layer.enable_linear)
_copy_weights(layer, ref)
x = torch.randn(2, 3, 224, 224)
with torch.no_grad():
torch.testing.assert_close(
layer.forward_native(x), ref(x), rtol=1e-4, atol=1e-4
)
def test_enable_linear(self):
layer = Conv2dLayer(
3, 768, kernel_size=14, stride=14, bias=True, disable_linear=False
)
ref = nn.Conv2d(3, 768, kernel_size=14, stride=14, bias=True)
self.assertTrue(layer.enable_linear)
_copy_weights(layer, ref)
x = torch.randn(1, 3, 224, 224)
with torch.no_grad():
torch.testing.assert_close(
layer.forward_native(x), ref(x), rtol=1e-4, atol=1e-4
)
def test_padding_valid(self):
layer = Conv2dLayer(3, 768, kernel_size=14, stride=14, padding="valid")
self.assertFalse(layer.enable_linear)
self.assertEqual(layer.padding, (0, 0))
def test_padding_same_disables_linear(self):
layer = Conv2dLayer(3, 64, kernel_size=3, stride=1, padding="same")
self.assertFalse(layer.enable_linear)
def test_non_matching_stride_disables_linear(self):
layer = Conv2dLayer(3, 64, kernel_size=3, stride=1, padding=1)
self.assertFalse(layer.enable_linear)
def test_groups_disable_linear(self):
layer = Conv2dLayer(4, 8, kernel_size=2, stride=2, groups=2)
self.assertFalse(layer.enable_linear)
def test_default_disables_linear(self):
layer = Conv2dLayer(3, 768, kernel_size=14, stride=14)
self.assertFalse(layer.enable_linear)
def test_dilation_disables_linear(self):
layer = Conv2dLayer(3, 64, kernel_size=3, stride=3, dilation=2)
self.assertFalse(layer.enable_linear)
def test_padding_mode_reflect(self):
layer = Conv2dLayer(
3, 64, kernel_size=3, stride=1, padding=1, padding_mode="reflect", bias=True
)
ref = nn.Conv2d(
3, 64, kernel_size=3, stride=1, padding=1, padding_mode="reflect", bias=True
)
self.assertFalse(layer.enable_linear)
_copy_weights(layer, ref)
x = torch.randn(1, 3, 16, 16)
with torch.no_grad():
torch.testing.assert_close(
layer.forward_native(x), ref(x), rtol=1e-4, atol=1e-4
)
def test_conv_path_with_padding(self):
layer = Conv2dLayer(3, 64, kernel_size=3, stride=1, padding=1, bias=True)
ref = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=True)
_copy_weights(layer, ref)
x = torch.randn(1, 3, 32, 32)
with torch.no_grad():
torch.testing.assert_close(
layer.forward_native(x), ref(x), rtol=1e-4, atol=1e-4
)
def test_mulmat_matches_conv(self):
layer = Conv2dLayer(
3, 768, kernel_size=14, stride=14, bias=True, disable_linear=False
)
self.assertTrue(layer.enable_linear)
x = torch.randn(2, 3, 224, 224)
with torch.no_grad():
torch.testing.assert_close(
layer._forward_mulmat(x),
layer._forward_conv(x),
rtol=1e-4,
atol=1e-4,
)
def test_forward_cuda_uses_mulmat_when_enabled(self):
layer = Conv2dLayer(
3, 64, kernel_size=4, stride=4, bias=False, disable_linear=False
)
self.assertTrue(layer.enable_linear)
x = torch.randn(1, 3, 16, 16)
with torch.no_grad():
torch.testing.assert_close(layer.forward_cuda(x), layer._forward_mulmat(x))
def test_forward_cuda_uses_conv_when_not_eligible(self):
layer = Conv2dLayer(3, 64, kernel_size=3, stride=1, padding=1, bias=False)
self.assertFalse(layer.enable_linear)
x = torch.randn(1, 3, 16, 16)
with torch.no_grad():
torch.testing.assert_close(layer.forward_cuda(x), layer._forward_conv(x))
def test_tuple_kernel_size(self):
layer = Conv2dLayer(
3,
768,
kernel_size=(14, 14),
stride=(14, 14),
bias=False,
disable_linear=False,
)
self.assertTrue(layer.enable_linear)
ref = nn.Conv2d(3, 768, kernel_size=(14, 14), stride=(14, 14), bias=False)
_copy_weights(layer, ref)
x = torch.randn(1, 3, 224, 224)
with torch.no_grad():
torch.testing.assert_close(
layer.forward_native(x), ref(x), rtol=1e-4, atol=1e-4
)
def test_output_shape(self):
layer = Conv2dLayer(3, 768, kernel_size=16, stride=16, bias=False)
x = torch.randn(4, 3, 224, 224)
out = layer.forward_native(x)
self.assertEqual(out.shape, (4, 768, 14, 14))
def test_no_bias_parameter(self):
layer = Conv2dLayer(3, 64, kernel_size=4, stride=4, bias=False)
self.assertIsNone(layer.bias)
class TestConvValidation(unittest.TestCase):
def test_in_channels_not_divisible_by_groups(self):
with self.assertRaises(ValueError):
Conv2dLayer(3, 64, kernel_size=3, stride=1, groups=2)
def test_out_channels_not_divisible_by_groups(self):
with self.assertRaises(ValueError):
Conv2dLayer(4, 6, kernel_size=3, stride=1, groups=4)
def test_invalid_padding_string(self):
with self.assertRaises(ValueError):
Conv2dLayer(3, 64, kernel_size=3, stride=1, padding="full")
def test_padding_same_with_stride(self):
with self.assertRaises(ValueError):
Conv2dLayer(3, 64, kernel_size=3, stride=2, padding="same")
def test_padding_same_with_non_zeros_padding_mode(self):
layer = Conv2dLayer(
3,
64,
kernel_size=3,
stride=1,
padding="same",
padding_mode="reflect",
bias=True,
)
ref = nn.Conv2d(
3,
64,
kernel_size=3,
stride=1,
padding="same",
padding_mode="reflect",
bias=True,
)
self.assertFalse(layer.enable_linear)
_copy_weights(layer, ref)
x = torch.randn(1, 3, 16, 16)
with torch.no_grad():
torch.testing.assert_close(
layer.forward_native(x), ref(x), rtol=1e-4, atol=1e-4
)
def test_invalid_padding_mode(self):
with self.assertRaises(ValueError):
Conv3dLayer(3, 64, kernel_size=3, stride=1, padding_mode="invalid")
def test_conv3d_in_channels_not_divisible_by_groups(self):
with self.assertRaises(ValueError):
Conv3dLayer(3, 64, kernel_size=3, stride=1, groups=2)
class TestConv3dLayer(unittest.TestCase):
def test_basic_temporal_patch_embedding(self):
layer = Conv3dLayer(
3, 1152, kernel_size=[2, 14, 14], stride=[2, 14, 14], bias=False
)
ref = nn.Conv3d(
3, 1152, kernel_size=[2, 14, 14], stride=[2, 14, 14], bias=False
)
self.assertTrue(layer.enable_linear)
_copy_weights(layer, ref)
x = torch.randn(1, 3, 2, 14, 14)
with torch.no_grad():
torch.testing.assert_close(
layer.forward_native(x), ref(x), rtol=1e-4, atol=1e-4
)
def test_with_bias(self):
layer = Conv3dLayer(
3, 1536, kernel_size=[2, 14, 14], stride=[2, 14, 14], bias=True
)
ref = nn.Conv3d(3, 1536, kernel_size=[2, 14, 14], stride=[2, 14, 14], bias=True)
self.assertTrue(layer.enable_linear)
_copy_weights(layer, ref)
x = torch.randn(4, 3, 2, 14, 14)
with torch.no_grad():
torch.testing.assert_close(
layer.forward_native(x), ref(x), rtol=1e-4, atol=1e-4
)
def test_mulmat_matches_conv(self):
layer = Conv3dLayer(
3, 1152, kernel_size=[2, 14, 14], stride=[2, 14, 14], bias=True
)
self.assertTrue(layer.enable_linear)
x = torch.randn(2, 3, 2, 14, 14)
with torch.no_grad():
torch.testing.assert_close(
layer._forward_mulmat(x),
layer._forward_conv(x),
rtol=1e-4,
atol=1e-4,
)
def test_non_matching_stride_disables_linear(self):
layer = Conv3dLayer(3, 64, kernel_size=3, stride=1, padding=1)
self.assertFalse(layer.enable_linear)
def test_dilation_disables_linear(self):
layer = Conv3dLayer(3, 64, kernel_size=3, stride=3, dilation=2)
self.assertFalse(layer.enable_linear)
def test_disable_linear(self):
layer = Conv3dLayer(
3,
1152,
kernel_size=[2, 14, 14],
stride=[2, 14, 14],
bias=False,
disable_linear=True,
)
self.assertFalse(layer.enable_linear)
ref = nn.Conv3d(
3, 1152, kernel_size=[2, 14, 14], stride=[2, 14, 14], bias=False
)
_copy_weights(layer, ref)
x = torch.randn(1, 3, 2, 14, 14)
with torch.no_grad():
torch.testing.assert_close(
layer.forward_native(x), ref(x), rtol=1e-4, atol=1e-4
)
def test_conv_path_with_padding(self):
layer = Conv3dLayer(3, 64, kernel_size=3, stride=1, padding=1, bias=True)
ref = nn.Conv3d(3, 64, kernel_size=3, stride=1, padding=1, bias=True)
_copy_weights(layer, ref)
x = torch.randn(1, 3, 4, 8, 8)
with torch.no_grad():
torch.testing.assert_close(
layer.forward_native(x), ref(x), rtol=1e-4, atol=1e-4
)
def test_output_shape(self):
layer = Conv3dLayer(
3, 1152, kernel_size=[2, 14, 14], stride=[2, 14, 14], bias=False
)
x = torch.randn(1, 3, 2, 14, 14)
out = layer.forward_native(x)
self.assertEqual(out.shape, (1, 1152, 1, 1, 1))
def test_batch_processing(self):
layer = Conv3dLayer(
3, 1536, kernel_size=[2, 14, 14], stride=[2, 14, 14], bias=True
)
ref = nn.Conv3d(3, 1536, kernel_size=[2, 14, 14], stride=[2, 14, 14], bias=True)
_copy_weights(layer, ref)
x = torch.randn(8, 3, 2, 14, 14)
with torch.no_grad():
torch.testing.assert_close(
layer.forward_native(x), ref(x), rtol=1e-4, atol=1e-4
)
def test_forward_native_uses_mulmat_when_eligible(self):
layer = Conv3dLayer(3, 128, kernel_size=[2, 4, 4], stride=[2, 4, 4], bias=True)
self.assertTrue(layer.enable_linear)
x = torch.randn(1, 3, 2, 4, 4)
with torch.no_grad():
torch.testing.assert_close(
layer.forward_native(x), layer._forward_mulmat(x)
)
def test_padding_valid(self):
layer = Conv3dLayer(
3, 64, kernel_size=[2, 4, 4], stride=[2, 4, 4], padding="valid"
)
self.assertTrue(layer.enable_linear)
self.assertEqual(layer.padding, (0, 0, 0))
def test_weight_shape(self):
layer = Conv3dLayer(
3, 1152, kernel_size=[2, 14, 14], stride=[2, 14, 14], bias=False
)
self.assertEqual(layer.weight.shape, (1152, 3, 2, 14, 14))
def test_glm4v_workflow(self):
"""GLM4V-style: 2D input -> reshape to 5D -> Conv3dLayer -> flatten."""
in_channels, temporal_patch_size, patch_size = 3, 2, 14
hidden_size = 1536
layer = Conv3dLayer(
in_channels,
hidden_size,
kernel_size=[temporal_patch_size, patch_size, patch_size],
stride=[temporal_patch_size, patch_size, patch_size],
bias=True,
)
ref = nn.Conv3d(
in_channels,
hidden_size,
kernel_size=[temporal_patch_size, patch_size, patch_size],
stride=[temporal_patch_size, patch_size, patch_size],
bias=True,
)
_copy_weights(layer, ref)
num_patches = 4
flat_dim = in_channels * temporal_patch_size * patch_size * patch_size
x_2d = torch.randn(num_patches, flat_dim)
x_5d = x_2d.view(-1, in_channels, temporal_patch_size, patch_size, patch_size)
with torch.no_grad():
torch.testing.assert_close(
layer.forward_native(x_5d).view(-1, hidden_size),
ref(x_5d).view(-1, hidden_size),
rtol=1e-4,
atol=1e-4,
)
if __name__ == "__main__":
unittest.main()
test/registered/unit/layers/test_mamba_state_scatter_triton.py
+358
View File
@@ -0,0 +1,358 @@
from sglang.test.ci.ci_register import register_cuda_ci
register_cuda_ci(est_time=10, suite="stage-b-test-small-1-gpu")
import os
import unittest
import torch
try:
from sglang.srt.layers.attention.mamba.mamba_state_scatter_triton import (
fused_mamba_state_scatter_with_mask,
)
_FUSED_IMPORT_ERROR = None
except Exception as e: # pragma: no cover
fused_mamba_state_scatter_with_mask = None
_FUSED_IMPORT_ERROR = e
def _dtype_from_str(name: str) -> torch.dtype:
mapping = {
"bfloat16": torch.bfloat16,
"float16": torch.float16,
"float32": torch.float32,
}
if name not in mapping:
raise ValueError(
f"Unsupported dtype string {name!r}. Supported: {sorted(mapping.keys())}"
)
return mapping[name]
def _ref_scatter(dst, src, dst_indices, src_indices, step_indices):
"""Reference implementation using PyTorch advanced indexing."""
# dst: [L, C, E]
# src: [L, S, D, E]
dst[:, dst_indices] = src[:, src_indices, step_indices].to(dst.dtype, copy=False)
def _ref_update_like(
ssm_states,
intermediate_ssm,
conv_states,
intermediate_conv,
*,
state_indices_tensor,
accepted_steps,
mamba_track_indices=None,
mamba_steps_to_track=None,
):
"""Reference implementation using PyTorch advanced indexing for correctness verification."""
request_number = accepted_steps.shape[0]
intermediate_state_indices = torch.arange(
request_number, dtype=torch.int32, device=accepted_steps.device
)
valid_mask = accepted_steps >= 0
dst_state_indices = state_indices_tensor[valid_mask].to(torch.int64)
src_state_indices = intermediate_state_indices[valid_mask].to(torch.int64)
last_steps = accepted_steps[valid_mask].to(torch.int64)
# Only scatter if there are valid indices (but don't early return -
# mamba_track_indices processing is independent)
if dst_state_indices.numel() > 0:
_ref_scatter(
ssm_states,
intermediate_ssm,
dst_state_indices,
src_state_indices,
last_steps,
)
_ref_scatter(
conv_states,
intermediate_conv,
dst_state_indices,
src_state_indices,
last_steps,
)
if mamba_track_indices is not None:
assert mamba_steps_to_track is not None
track_mask = mamba_steps_to_track >= 0
if not track_mask.any():
return
dst_track_indices = mamba_track_indices[track_mask].to(torch.int64)
src_track_indices = intermediate_state_indices[track_mask].to(torch.int64)
track_steps = mamba_steps_to_track[track_mask].to(torch.int64)
_ref_scatter(
ssm_states,
intermediate_ssm,
dst_track_indices,
src_track_indices,
track_steps,
)
_ref_scatter(
conv_states,
intermediate_conv,
dst_track_indices,
src_track_indices,
track_steps,
)
def _fused_update_like(
ssm_states,
intermediate_ssm,
conv_states,
intermediate_conv,
*,
state_indices_tensor,
accepted_steps,
mamba_track_indices=None,
mamba_steps_to_track=None,
):
"""Matches the fully fused logic that avoids index_select and nonzero calls."""
# Use fully fused kernel that handles masking internally
fused_mamba_state_scatter_with_mask(
ssm_states,
intermediate_ssm,
state_indices_tensor,
accepted_steps,
)
fused_mamba_state_scatter_with_mask(
conv_states,
intermediate_conv,
state_indices_tensor,
accepted_steps,
)
if mamba_track_indices is not None:
assert mamba_steps_to_track is not None
fused_mamba_state_scatter_with_mask(
ssm_states,
intermediate_ssm,
mamba_track_indices,
mamba_steps_to_track,
)
fused_mamba_state_scatter_with_mask(
conv_states,
intermediate_conv,
mamba_track_indices,
mamba_steps_to_track,
)
def _time_cuda_ms(fn, iters=50, warmup=10):
"""Measure average CUDA time (ms) using CUDA events."""
for _ in range(warmup):
fn()
torch.cuda.synchronize()
start = torch.cuda.Event(enable_timing=True)
end = torch.cuda.Event(enable_timing=True)
start.record()
for _ in range(iters):
fn()
end.record()
torch.cuda.synchronize()
return start.elapsed_time(end) / iters
class TestMambaStateScatterCorrectness(unittest.TestCase):
@unittest.skipUnless(torch.cuda.is_available(), "CUDA is required for this test.")
def test_fused_matches_reference(self):
"""Test that fused_mamba_state_scatter_with_mask matches the reference."""
if fused_mamba_state_scatter_with_mask is None:
self.skipTest(
f"fused_mamba_state_scatter_with_mask import failed: {_FUSED_IMPORT_ERROR}"
)
torch.manual_seed(42)
device = torch.device("cuda")
# Keep sizes moderate so this test is quick.
L = 8
B = 32
C = 49
D = 5
ssm_elems = 1024
conv_elems = 512
ssm_states0 = torch.randn(
(L, C, ssm_elems), device=device, dtype=torch.bfloat16
)
conv_states0 = torch.randn(
(L, C, conv_elems), device=device, dtype=torch.bfloat16
)
intermediate_ssm = torch.randn(
(L, B, D, ssm_elems), device=device, dtype=torch.bfloat16
)
intermediate_conv = torch.randn(
(L, B, D, conv_elems), device=device, dtype=torch.bfloat16
)
# unique cache lines (no duplicates) to avoid nondeterministic write order
state_indices_tensor = torch.randperm(C, device=device, dtype=torch.int64)[
:B
].to(torch.int32)
accepted_steps = torch.randint(0, D, (B,), device=device, dtype=torch.int64)
# set ~10% invalid
invalid = torch.rand((B,), device=device) < 0.1
accepted_steps[invalid] = -1
# Optional track update
mamba_track_indices = torch.randperm(C, device=device, dtype=torch.int64)[:B]
mamba_steps_to_track = torch.randint(
0, D, (B,), device=device, dtype=torch.int64
)
track_invalid = torch.rand((B,), device=device) < 0.7
mamba_steps_to_track[track_invalid] = -1
ssm_ref = ssm_states0.clone()
conv_ref = conv_states0.clone()
ssm_fused = ssm_states0.clone()
conv_fused = conv_states0.clone()
_ref_update_like(
ssm_ref,
intermediate_ssm,
conv_ref,
intermediate_conv,
state_indices_tensor=state_indices_tensor,
accepted_steps=accepted_steps,
mamba_track_indices=mamba_track_indices,
mamba_steps_to_track=mamba_steps_to_track,
)
_fused_update_like(
ssm_fused,
intermediate_ssm,
conv_fused,
intermediate_conv,
state_indices_tensor=state_indices_tensor,
accepted_steps=accepted_steps,
mamba_track_indices=mamba_track_indices,
mamba_steps_to_track=mamba_steps_to_track,
)
torch.testing.assert_close(ssm_fused, ssm_ref)
torch.testing.assert_close(conv_fused, conv_ref)
class TestMambaStateScatterPerf(unittest.TestCase):
@unittest.skipUnless(torch.cuda.is_available(), "CUDA is required for this test.")
def test_perf_report_old_vs_fused(self):
"""Optional microbenchmark comparing baseline vs fused kernel.
Enable with: SGLANG_RUN_MAMBA_SCATTER_PERF_TEST=1
"""
if os.environ.get("SGLANG_RUN_MAMBA_SCATTER_PERF_TEST", "0") != "1":
self.skipTest("Set SGLANG_RUN_MAMBA_SCATTER_PERF_TEST=1 to run perf test.")
if fused_mamba_state_scatter_with_mask is None:
self.skipTest(
f"fused_mamba_state_scatter_with_mask import failed: {_FUSED_IMPORT_ERROR}"
)
torch.manual_seed(0)
device = torch.device("cuda")
# Parameterize sizes via env vars so we can match a real model more closely.
L = int(os.environ.get("SGLANG_MAMBA_SCATTER_LAYERS", "32"))
B = int(os.environ.get("SGLANG_MAMBA_SCATTER_BATCH", "48"))
C = int(os.environ.get("SGLANG_MAMBA_SCATTER_CACHE", "49"))
D = int(os.environ.get("SGLANG_MAMBA_SCATTER_DRAFT_TOKENS", "5"))
ssm_elems = int(os.environ.get("SGLANG_MAMBA_SCATTER_SSM_ELEMS", "4096"))
conv_elems = int(os.environ.get("SGLANG_MAMBA_SCATTER_CONV_ELEMS", "512"))
invalid_ratio = float(
os.environ.get("SGLANG_MAMBA_SCATTER_INVALID_RATIO", "0.0")
)
track_ratio = float(os.environ.get("SGLANG_MAMBA_SCATTER_TRACK_RATIO", "0.0"))
ssm_dtype = _dtype_from_str(
os.environ.get("SGLANG_MAMBA_SCATTER_SSM_DTYPE", "bfloat16")
)
conv_dtype = _dtype_from_str(
os.environ.get("SGLANG_MAMBA_SCATTER_CONV_DTYPE", "bfloat16")
)
# Use zeros for dst so each iteration overwrites the same memory.
ssm_states = torch.zeros((L, C, ssm_elems), device=device, dtype=ssm_dtype)
conv_states = torch.zeros((L, C, conv_elems), device=device, dtype=conv_dtype)
intermediate_ssm = torch.randn(
(L, B, D, ssm_elems), device=device, dtype=ssm_dtype
)
intermediate_conv = torch.randn(
(L, B, D, conv_elems), device=device, dtype=conv_dtype
)
state_indices_tensor = torch.randperm(C, device=device, dtype=torch.int64)[
:B
].to(torch.int32)
accepted_steps = torch.randint(0, D, (B,), device=device, dtype=torch.int64)
if invalid_ratio > 0:
invalid = torch.rand((B,), device=device) < invalid_ratio
accepted_steps[invalid] = -1
mamba_track_indices = None
mamba_steps_to_track = None
if track_ratio > 0:
mamba_track_indices = torch.randperm(C, device=device, dtype=torch.int64)[
:B
]
mamba_steps_to_track = torch.randint(
0, D, (B,), device=device, dtype=torch.int64
)
track_invalid = torch.rand((B,), device=device) >= track_ratio
mamba_steps_to_track[track_invalid] = -1
def ref_fn():
_ref_update_like(
ssm_states,
intermediate_ssm,
conv_states,
intermediate_conv,
state_indices_tensor=state_indices_tensor,
accepted_steps=accepted_steps,
mamba_track_indices=mamba_track_indices,
mamba_steps_to_track=mamba_steps_to_track,
)
def fused_fn():
_fused_update_like(
ssm_states,
intermediate_ssm,
conv_states,
intermediate_conv,
state_indices_tensor=state_indices_tensor,
accepted_steps=accepted_steps,
mamba_track_indices=mamba_track_indices,
mamba_steps_to_track=mamba_steps_to_track,
)
# Warm up JIT compilation for triton kernels (and caches for torch indexing)
ref_fn()
fused_fn()
torch.cuda.synchronize()
ref_ms = _time_cuda_ms(ref_fn)
fused_ms = _time_cuda_ms(fused_fn)
num_valid = int((accepted_steps >= 0).sum().item())
ratio = fused_ms / ref_ms if ref_ms > 0 else float("inf")
speedup = ref_ms / fused_ms if fused_ms > 0 else float("inf")
# Print a concise report
print(
"\n[MambaStateScatterPerf]\n"
f" shapes: L={L} B={B} C={C} D={D} ssm_elems={ssm_elems} conv_elems={conv_elems}\n"
f" dtypes: ssm={ssm_dtype} conv={conv_dtype}\n"
f" valid: {num_valid}/{B} invalid_ratio={invalid_ratio} track_ratio={track_ratio}\n"
f" ref_total_ms (baseline): {ref_ms:.4f}\n"
f" fused_total_ms: {fused_ms:.4f} (ratio={ratio:.3f}x, speedup={speedup:.2f}x)\n"
)
if __name__ == "__main__": # pragma: no cover
unittest.main()