Skip to content

Commit 866096d

Browse files
LLLLKKKKLLLLKKKK
committed
test: rm bench in deepgemm test
1 parent f8fe77e commit 866096d

File tree

1 file changed

+127
-127
lines changed

1 file changed

+127
-127
lines changed

‎rtp_llm/models_py/utils/test/deepgemm_test.py‎

Lines changed: 127 additions & 127 deletions
Original file line numberDiff line numberDiff line change
@@ -315,28 +315,28 @@ def test_fp8_gemm_nt(self) -> None:
315315
f"{m=}, {n=}, {k=}, {kernel_opt}, {accumulate=}, {out_dtype=}, "
316316
f"{diff:.5f}, alias={test_alias}"
317317
)
318-
a, b, c, d, ref_d = generate_normal(
319-
m, n, k, accumulate, out_dtype, use_ue8m0=use_ue8m0
320-
)
321-
322-
# Test launch overhead
323-
launch_start_t = time.time_ns()
324-
fp8_gemm_nt(a, b, d, c=c, disable_ue8m0_cast=disable_ue8m0_cast)
325-
launch_end_t = time.time_ns()
326-
torch.cuda.synchronize()
327-
328-
# noinspection PyShadowingNames
329-
def test_func():
330-
fp8_gemm_nt(a, b, d, c=c, disable_ue8m0_cast=disable_ue8m0_cast)
331-
332-
t = bench_kineto(test_func, "fp8_gemm", suppress_kineto_output=True)
333-
print(
334-
f" > Perf (m={m:5}, n={n:5}, k={k:5}, {kernel_opt}, {out_opt}, {acc_opt}): "
335-
f"launch {(launch_end_t - launch_start_t) / 1e3:4.0f} us | {t * 1e6:4.0f} us | "
336-
f"{2 * m * n * k / t / 1e12:4.0f} TFLOPS | "
337-
f"{(count_bytes(a, b, d) + count_bytes(c) * int(accumulate)) / 1e9 / t:4.0f} GB/s",
338-
flush=True,
339-
)
318+
# a, b, c, d, ref_d = generate_normal(
319+
# m, n, k, accumulate, out_dtype, use_ue8m0=use_ue8m0
320+
# )
321+
322+
# # Test launch overhead
323+
# launch_start_t = time.time_ns()
324+
# fp8_gemm_nt(a, b, d, c=c, disable_ue8m0_cast=disable_ue8m0_cast)
325+
# launch_end_t = time.time_ns()
326+
# torch.cuda.synchronize()
327+
328+
# # noinspection PyShadowingNames
329+
# def test_func():
330+
# fp8_gemm_nt(a, b, d, c=c, disable_ue8m0_cast=disable_ue8m0_cast)
331+
332+
# t = bench_kineto(test_func, "fp8_gemm", suppress_kineto_output=True)
333+
# print(
334+
# f" > Perf (m={m:5}, n={n:5}, k={k:5}, {kernel_opt}, {out_opt}, {acc_opt}): "
335+
# f"launch {(launch_end_t - launch_start_t) / 1e3:4.0f} us | {t * 1e6:4.0f} us | "
336+
# f"{2 * m * n * k / t / 1e12:4.0f} TFLOPS | "
337+
# f"{(count_bytes(a, b, d) + count_bytes(c) * int(accumulate)) / 1e9 / t:4.0f} GB/s",
338+
# flush=True,
339+
# )
340340
print(flush=True)
341341

342342
def test_m_grouped_fp8_gemm_nt_contiguous(self) -> None:
@@ -367,24 +367,24 @@ def test_m_grouped_fp8_gemm_nt_contiguous(self) -> None:
367367
assert (
368368
diff < 0.001
369369
), f"{m=}, {n=}, {k=}, {kernel_opt}, {diff:.5f}, alias={test_alias}"
370-
m, a, b, m_indices, d, ref_d = generate_m_grouped_contiguous(
371-
num_groups, expected_m_per_group, n, k, use_ue8m0=use_ue8m0
372-
)
373-
374-
# noinspection PyShadowingNames
375-
def test_func():
376-
m_grouped_fp8_gemm_nt_contiguous(
377-
a, b, d, m_indices, disable_ue8m0_cast=disable_ue8m0_cast
378-
)
379-
380-
t = bench_kineto(test_func, "fp8_gemm", suppress_kineto_output=True)
381-
print(
382-
f" > Perf ({num_groups=}, m={m:5}, n={n:5}, k={k:5}, {kernel_opt}): "
383-
f"{t * 1e6:4.0f} us | "
384-
f"{2 * m * n * k / t / 1e12:4.0f} TFLOPS | "
385-
f"{count_bytes(a, b, d) / 1e9 / t:4.0f} GB/s",
386-
flush=True,
387-
)
370+
# m, a, b, m_indices, d, ref_d = generate_m_grouped_contiguous(
371+
# num_groups, expected_m_per_group, n, k, use_ue8m0=use_ue8m0
372+
# )
373+
374+
# # noinspection PyShadowingNames
375+
# def test_func():
376+
# m_grouped_fp8_gemm_nt_contiguous(
377+
# a, b, d, m_indices, disable_ue8m0_cast=disable_ue8m0_cast
378+
# )
379+
380+
# t = bench_kineto(test_func, "fp8_gemm", suppress_kineto_output=True)
381+
# print(
382+
# f" > Perf ({num_groups=}, m={m:5}, n={n:5}, k={k:5}, {kernel_opt}): "
383+
# f"{t * 1e6:4.0f} us | "
384+
# f"{2 * m * n * k / t / 1e12:4.0f} TFLOPS | "
385+
# f"{count_bytes(a, b, d) / 1e9 / t:4.0f} GB/s",
386+
# flush=True,
387+
# )
388388
print(flush=True)
389389

390390
def test_m_grouped_fp8_gemm_nt_masked(self) -> None:
@@ -424,32 +424,32 @@ def test_m_grouped_fp8_gemm_nt_masked(self) -> None:
424424
diff < 0.001
425425
), f"{max_m=}, {n=}, {k=}, {j=}, masked_m={masked_m[j]}, {kernel_opt}, {num_groups=}, {diff:.5f}"
426426

427-
# Construct full cases
428-
a, b, masked_m, d, ref_d = generate_m_grouped_masked(
429-
num_groups, max_m, expected_m_per_group, n, k, use_ue8m0=use_ue8m0
430-
)
431-
432-
# noinspection PyShadowingNames
433-
def test_func():
434-
m_grouped_fp8_gemm_nt_masked(
435-
a,
436-
b,
437-
d,
438-
masked_m,
439-
expected_m_per_group,
440-
disable_ue8m0_cast=disable_ue8m0_cast,
441-
)
442-
443-
# Test performance with fixed shapes
444-
valid_m = masked_m.sum().item()
445-
t = bench_kineto(test_func, "fp8_gemm", suppress_kineto_output=True)
446-
print(
447-
f" > Perf ({num_groups=}, expected_m_per_group={expected_m_per_group:4}, n={n:4}, k={k:4}, {kernel_opt}): "
448-
f"{t * 1e6:4.0f} us | "
449-
f"{2 * valid_m * n * k / t / 1e12:4.0f} TFLOPS | "
450-
f"{(count_bytes(a, d) * valid_m / (max_m * num_groups) + count_bytes(b)) / 1e9 / t:4.0f} GB/s",
451-
flush=True,
452-
)
427+
# # Construct full cases
428+
# a, b, masked_m, d, ref_d = generate_m_grouped_masked(
429+
# num_groups, max_m, expected_m_per_group, n, k, use_ue8m0=use_ue8m0
430+
# )
431+
432+
# # noinspection PyShadowingNames
433+
# def test_func():
434+
# m_grouped_fp8_gemm_nt_masked(
435+
# a,
436+
# b,
437+
# d,
438+
# masked_m,
439+
# expected_m_per_group,
440+
# disable_ue8m0_cast=disable_ue8m0_cast,
441+
# )
442+
443+
# # Test performance with fixed shapes
444+
# valid_m = masked_m.sum().item()
445+
# t = bench_kineto(test_func, "fp8_gemm", suppress_kineto_output=True)
446+
# print(
447+
# f" > Perf ({num_groups=}, expected_m_per_group={expected_m_per_group:4}, n={n:4}, k={k:4}, {kernel_opt}): "
448+
# f"{t * 1e6:4.0f} us | "
449+
# f"{2 * valid_m * n * k / t / 1e12:4.0f} TFLOPS | "
450+
# f"{(count_bytes(a, d) * valid_m / (max_m * num_groups) + count_bytes(b)) / 1e9 / t:4.0f} GB/s",
451+
# flush=True,
452+
# )
453453
print(flush=True)
454454

455455
def test_bf16_gemm_nt(self) -> None:
@@ -470,32 +470,32 @@ def test_bf16_gemm_nt(self) -> None:
470470
f"{m=}, {n=}, {k=}, {accumulate=}, {out_dtype=}, "
471471
f"{diff:.5f}, alias={test_alias}"
472472
)
473-
a, b, c, d, ref_d = generate_normal(
474-
m, n, k, accumulate, out_dtype, use_bf16=True
475-
)
476-
477-
cublas_t = 0
478-
t = bench_kineto(
479-
lambda: bf16_gemm_nt(a, b, d, c=c),
480-
"bf16_gemm",
481-
suppress_kineto_output=True,
482-
)
483-
if accumulate == 0 and out_dtype == torch.bfloat16:
484-
# noinspection PyBroadException
485-
try:
486-
cublas_t = bench_kineto(
487-
lambda: a @ b.T, "nvjet", suppress_kineto_output=True
488-
)
489-
except Exception:
490-
pass
491-
print(
492-
f" > Perf (m={m:5}, n={n:5}, k={k:5}, {out_opt}, {acc_opt}): "
493-
f"{t * 1e6:4.0f} us | "
494-
f"{2 * m * n * k / t / 1e12:4.0f} TFLOPS | "
495-
f"{(count_bytes(a, b, d) + count_bytes(c) * int(accumulate)) / 1e9 / t:4.0f} GB/s | "
496-
f"{cublas_t / t:.2f}x cuBLAS",
497-
flush=True,
498-
)
473+
# a, b, c, d, ref_d = generate_normal(
474+
# m, n, k, accumulate, out_dtype, use_bf16=True
475+
# )
476+
477+
# cublas_t = 0
478+
# t = bench_kineto(
479+
# lambda: bf16_gemm_nt(a, b, d, c=c),
480+
# "bf16_gemm",
481+
# suppress_kineto_output=True,
482+
# )
483+
# if accumulate == 0 and out_dtype == torch.bfloat16:
484+
# # noinspection PyBroadException
485+
# try:
486+
# cublas_t = bench_kineto(
487+
# lambda: a @ b.T, "nvjet", suppress_kineto_output=True
488+
# )
489+
# except Exception:
490+
# pass
491+
# print(
492+
# f" > Perf (m={m:5}, n={n:5}, k={k:5}, {out_opt}, {acc_opt}): "
493+
# f"{t * 1e6:4.0f} us | "
494+
# f"{2 * m * n * k / t / 1e12:4.0f} TFLOPS | "
495+
# f"{(count_bytes(a, b, d) + count_bytes(c) * int(accumulate)) / 1e9 / t:4.0f} GB/s | "
496+
# f"{cublas_t / t:.2f}x cuBLAS",
497+
# flush=True,
498+
# )
499499
print(flush=True)
500500

501501
def test_m_grouped_bf16_gemm_nt_contiguous(self) -> None:
@@ -514,22 +514,22 @@ def test_m_grouped_bf16_gemm_nt_contiguous(self) -> None:
514514
d = torch.where((m_indices == -1).unsqueeze(1), torch.zeros_like(d), d)
515515
diff = calc_diff(d, ref_d)
516516
assert diff < 0.001, f"{m=}, {n=}, {k=}, {diff:.5f}, alias={test_alias}"
517-
m, a, b, m_indices, d, ref_d = generate_m_grouped_contiguous(
518-
num_groups, expected_m_per_group, n, k, use_bf16=True
519-
)
520-
521-
# noinspection PyShadowingNames
522-
def test_func():
523-
m_grouped_bf16_gemm_nt_contiguous(a, b, d, m_indices)
524-
525-
t = bench_kineto(test_func, "bf16_gemm", suppress_kineto_output=True)
526-
print(
527-
f" > Perf ({num_groups=}, m={m:5}, n={n:5}, k={k:5}): "
528-
f"{t * 1e6:4.0f} us | "
529-
f"{2 * m * n * k / t / 1e12:4.0f} TFLOPS | "
530-
f"{count_bytes(a, b, d) / 1e9 / t:4.0f} GB/s",
531-
flush=True,
532-
)
517+
# m, a, b, m_indices, d, ref_d = generate_m_grouped_contiguous(
518+
# num_groups, expected_m_per_group, n, k, use_bf16=True
519+
# )
520+
521+
# # noinspection PyShadowingNames
522+
# def test_func():
523+
# m_grouped_bf16_gemm_nt_contiguous(a, b, d, m_indices)
524+
525+
# t = bench_kineto(test_func, "bf16_gemm", suppress_kineto_output=True)
526+
# print(
527+
# f" > Perf ({num_groups=}, m={m:5}, n={n:5}, k={k:5}): "
528+
# f"{t * 1e6:4.0f} us | "
529+
# f"{2 * m * n * k / t / 1e12:4.0f} TFLOPS | "
530+
# f"{count_bytes(a, b, d) / 1e9 / t:4.0f} GB/s",
531+
# flush=True,
532+
# )
533533
print(flush=True)
534534

535535
def test_m_grouped_bf16_gemm_nt_masked(self) -> None:
@@ -558,25 +558,25 @@ def test_m_grouped_bf16_gemm_nt_masked(self) -> None:
558558
diff < 0.001
559559
), f"{max_m=}, {n=}, {k=}, {j=}, masked_m={masked_m[j]}, {num_groups=}, {diff:.5f}"
560560

561-
# Construct full cases
562-
a, b, masked_m, d, ref_d = generate_m_grouped_masked(
563-
num_groups, max_m, expected_m_per_group, n, k, use_bf16=True
564-
)
565-
566-
# noinspection PyShadowingNames
567-
def test_func():
568-
m_grouped_bf16_gemm_nt_masked(a, b, d, masked_m, expected_m_per_group)
569-
570-
# Test performance with fixed shapes
571-
valid_m = masked_m.sum().item()
572-
t = bench_kineto(test_func, "bf16_gemm", suppress_kineto_output=True)
573-
print(
574-
f" > Perf ({num_groups=}, expected_m_per_group={expected_m_per_group:4}, n={n:4}, k={k:4}): "
575-
f"{t * 1e6:4.0f} us | "
576-
f"{2 * valid_m * n * k / t / 1e12:4.0f} TFLOPS | "
577-
f"{(count_bytes(a, d) * valid_m / (max_m * num_groups) + count_bytes(b)) / 1e9 / t:4.0f} GB/s",
578-
flush=True,
579-
)
561+
# # Construct full cases
562+
# a, b, masked_m, d, ref_d = generate_m_grouped_masked(
563+
# num_groups, max_m, expected_m_per_group, n, k, use_bf16=True
564+
# )
565+
566+
# # noinspection PyShadowingNames
567+
# def test_func():
568+
# m_grouped_bf16_gemm_nt_masked(a, b, d, masked_m, expected_m_per_group)
569+
570+
# # Test performance with fixed shapes
571+
# valid_m = masked_m.sum().item()
572+
# t = bench_kineto(test_func, "bf16_gemm", suppress_kineto_output=True)
573+
# print(
574+
# f" > Perf ({num_groups=}, expected_m_per_group={expected_m_per_group:4}, n={n:4}, k={k:4}): "
575+
# f"{t * 1e6:4.0f} us | "
576+
# f"{2 * valid_m * n * k / t / 1e12:4.0f} TFLOPS | "
577+
# f"{(count_bytes(a, d) * valid_m / (max_m * num_groups) + count_bytes(b)) / 1e9 / t:4.0f} GB/s",
578+
# flush=True,
579+
# )
580580
print(flush=True)
581581

582582

0 commit comments

Comments
 (0)