Skip to content

Fix large FP8 and MXFP8 performance repros #387

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
hwanseoc wants to merge 2 commits into from
Closed

Fix large FP8 and MXFP8 performance repros #387

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4574c55
Skip FP8 reference checks in perf mode
hwanseoc Jun 11, 2026
b3b0395
Fix large MXFP8 performance repros
hwanseoc Jun 29, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
158 changes: 85 additions & 73 deletions test/python/sdpa/fp8.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -311,6 +311,7 @@ def create_paged_container_and_block_table(tensor, block_size):
def exec_sdpa_fp8(cfg, request, cudnn_handle):
if request.config.option.dryrun:
pytest.skip("dryrun")
perf = request.config.getoption("--perf")

cudnn_version = LooseVersion(cudnn.backend_version_string())
if cudnn_version < "9.14.0":
Expand Down Expand Up @@ -420,12 +421,15 @@ def exec_sdpa_fp8(cfg, request, cudnn_handle):
else:
padding = None

o_ref, stats_ref, o_amax = compute_ref(q_fp8, k_fp8, v_fp8, attn_scale=attn_scale,
q_descale=q_descale_gpu, k_descale=k_descale_gpu, v_descale=v_descale_gpu,
s_scale=s_scale_gpu, s_descale=s_descale_gpu, torch_itype=torch_itype,
torch_otype=torch_otype, padding=padding,
left_bound=left_bound, right_bound=right_bound, diag_align=diag_align,
sink_token=sink_token_gpu, rescale_threshold=rescale_threshold)
if perf:
o_amax = 1.0
else:
o_ref, stats_ref, o_amax = compute_ref(q_fp8, k_fp8, v_fp8, attn_scale=attn_scale,
q_descale=q_descale_gpu, k_descale=k_descale_gpu, v_descale=v_descale_gpu,
s_scale=s_scale_gpu, s_descale=s_descale_gpu, torch_itype=torch_itype,
torch_otype=torch_otype, padding=padding,
left_bound=left_bound, right_bound=right_bound, diag_align=diag_align,
sink_token=sink_token_gpu, rescale_threshold=rescale_threshold)

o_scale_gpu = torch.tensor([get_fp8_scale_factor(o_amax, torch_otype)], dtype=torch.float, device="cuda")

Expand Down Expand Up @@ -497,29 +501,30 @@ def exec_sdpa_fp8(cfg, request, cudnn_handle):
variant_pack[int(GraphFwdUid.sink_token)] = sink_token_gpu

workspace = torch.empty(graph_fwd.get_workspace_size(), dtype=torch.uint8, device="cuda")
if request.config.getoption("--perf"):
if perf:
times_ms = time_execution(graph_fwd.execute, variant_pack, workspace, cudnn_handle)
print(f"@@@@ FP8 Fwd graph_fwd.execute avg_time_ms={times_ms.mean().item():.3f}")
profile_execution(graph_fwd.execute, variant_pack, workspace, cudnn_handle)
graph_fwd.execute(variant_pack, workspace, handle=cudnn_handle)
torch.cuda.synchronize()

# Compare forward output
if is_ragged:
o_ref_comp = convert_uniform_to_packed(torch.einsum("bshd->bhsd", o_ref), seq_len_q_ref, max_t_q)
else:
o_ref_comp = o_ref
if not perf:
if is_ragged:
o_ref_comp = convert_uniform_to_packed(torch.einsum("bshd->bhsd", o_ref), seq_len_q_ref, max_t_q)
else:
o_ref_comp = o_ref

o_gpu_float = o_gpu.detach().float() * get_fp8_descale_factor(o_amax, torch_otype)
o_ref_float = o_ref_comp.detach().float() * get_fp8_descale_factor(o_amax, torch_otype)
o_gpu_float = o_gpu.detach().float() * get_fp8_descale_factor(o_amax, torch_otype)
o_ref_float = o_ref_comp.detach().float() * get_fp8_descale_factor(o_amax, torch_otype)

if is_ragged:
t_idx = sum(seq_len_q_list)
o_gpu_float[t_idx:] = 0
o_ref_float[t_idx:] = 0
if is_ragged:
t_idx = sum(seq_len_q_list)
o_gpu_float[t_idx:] = 0
o_ref_float[t_idx:] = 0

atol, rtol = 0.08, 0.2
torch.testing.assert_close(o_gpu_float, o_ref_float, atol=atol, rtol=rtol)
atol, rtol = 0.08, 0.2
torch.testing.assert_close(o_gpu_float, o_ref_float, atol=atol, rtol=rtol)

# Backward pass
if not cfg.is_infer:
Expand All @@ -530,30 +535,36 @@ def exec_sdpa_fp8(cfg, request, cudnn_handle):
o_descale_gpu = torch.tensor([get_fp8_descale_factor(o_amax, torch_otype)], dtype=torch.float, device="cuda")
dO_descale_gpu = torch.tensor([get_fp8_descale_factor(dO_amax, torch_itype)], dtype=torch.float, device="cuda")

# Get unpacked BSHD references for backward
if is_ragged:
q_ref_bwd = torch.einsum("bhsd->bshd", convert_packed_to_uniform(q_gpu, seq_len_q_ref, s_qo))
k_ref_bwd = torch.einsum("bhsd->bshd", convert_packed_to_uniform(k_gpu, seq_len_kv_ref, s_kv))
v_ref_bwd = torch.einsum("bhsd->bshd", convert_packed_to_uniform(v_gpu, seq_len_kv_ref, s_kv))
o_ref_bwd = torch.einsum("bhsd->bshd", convert_packed_to_uniform(o_gpu, seq_len_q_ref, s_qo))
dO_ref_bwd = dO_fp8
if perf:
dP_amax = 1.0
dQ_amax = 1.0
dK_amax = 1.0
dV_amax = 1.0
else:
q_ref_bwd = q_gpu
k_ref_bwd = k_gpu
v_ref_bwd = v_gpu
o_ref_bwd = o_gpu
dO_ref_bwd = dO_fp8

padding_bwd = (seq_len_q_ref, seq_len_kv_ref) if is_ragged else None
dQ_ref, dK_ref, dV_ref, dSink_token_ref, dP_amax, dQ_amax, dK_amax, dV_amax = compute_ref_backward(
q_ref_bwd, k_ref_bwd, v_ref_bwd, o_ref_bwd, dO_ref_bwd, attn_scale=attn_scale,
q_descale=q_descale_gpu, k_descale=k_descale_gpu, v_descale=v_descale_gpu,
s_scale=s_scale_gpu, s_descale=s_descale_gpu, torch_itype=torch_itype,
o_descale=o_descale_gpu, dO_descale=dO_descale_gpu,
torch_otype=torch_otype, padding=padding_bwd,
left_bound=left_bound, right_bound=right_bound, diag_align=diag_align,
sink_token=sink_token_gpu
)
# Get unpacked BSHD references for backward
if is_ragged:
q_ref_bwd = torch.einsum("bhsd->bshd", convert_packed_to_uniform(q_gpu, seq_len_q_ref, s_qo))
k_ref_bwd = torch.einsum("bhsd->bshd", convert_packed_to_uniform(k_gpu, seq_len_kv_ref, s_kv))
v_ref_bwd = torch.einsum("bhsd->bshd", convert_packed_to_uniform(v_gpu, seq_len_kv_ref, s_kv))
o_ref_bwd = torch.einsum("bhsd->bshd", convert_packed_to_uniform(o_gpu, seq_len_q_ref, s_qo))
dO_ref_bwd = dO_fp8
else:
q_ref_bwd = q_gpu
k_ref_bwd = k_gpu
v_ref_bwd = v_gpu
o_ref_bwd = o_gpu
dO_ref_bwd = dO_fp8

padding_bwd = (seq_len_q_ref, seq_len_kv_ref) if is_ragged else None
dQ_ref, dK_ref, dV_ref, dSink_token_ref, dP_amax, dQ_amax, dK_amax, dV_amax = compute_ref_backward(
q_ref_bwd, k_ref_bwd, v_ref_bwd, o_ref_bwd, dO_ref_bwd, attn_scale=attn_scale,
q_descale=q_descale_gpu, k_descale=k_descale_gpu, v_descale=v_descale_gpu,
s_scale=s_scale_gpu, s_descale=s_descale_gpu, torch_itype=torch_itype,
o_descale=o_descale_gpu, dO_descale=dO_descale_gpu,
torch_otype=torch_otype, padding=padding_bwd,
left_bound=left_bound, right_bound=right_bound, diag_align=diag_align,
sink_token=sink_token_gpu
)

dP_descale_gpu = torch.tensor([get_fp8_descale_factor(dP_amax, torch_itype)], dtype=torch.float, device="cuda")
dQ_scale_gpu = torch.tensor([get_fp8_scale_factor(dQ_amax, torch_otype)], dtype=torch.float, device="cuda")
Expand Down Expand Up @@ -623,7 +634,7 @@ def exec_sdpa_fp8(cfg, request, cudnn_handle):
variant_pack_bwd[int(GraphBwdUid.dSink_token)] = dSink_token_gpu

workspace_bwd = torch.empty(graph_bwd.get_workspace_size(), dtype=torch.uint8, device="cuda")
if request.config.getoption("--perf"):
if perf:
times_ms = time_execution(graph_bwd.execute, variant_pack_bwd, workspace_bwd, cudnn_handle)
print(f"@@@@ FP8 Bwd graph.execute avg_time_ms={times_ms.mean().item():.3f}")
profile_execution(graph_bwd.execute, variant_pack_bwd, workspace_bwd, cudnn_handle)
Expand Down Expand Up @@ -652,36 +663,37 @@ def exec_sdpa_fp8(cfg, request, cudnn_handle):
pytest.fail("determinism check failed", pytrace=False)
print("@@@@ Determinism check: PASSED, dQ, dK, dV bitwise match between runs.")

if is_ragged:
dQ_ref = convert_uniform_to_packed(torch.einsum("bshd->bhsd", dQ_ref), seq_len_q_ref, max_t_q)
dK_ref = convert_uniform_to_packed(torch.einsum("bshd->bhsd", dK_ref), seq_len_kv_ref, max_t_kv)
dV_ref = convert_uniform_to_packed(torch.einsum("bshd->bhsd", dV_ref), seq_len_kv_ref, max_t_kv)

dQ_out = dQ_gpu.detach().float() * get_fp8_descale_factor(dQ_amax, torch_otype)
dK_out = dK_gpu.detach().float() * get_fp8_descale_factor(dK_amax, torch_otype)
dV_out = dV_gpu.detach().float() * get_fp8_descale_factor(dV_amax, torch_otype)

dQ_ref_float = dQ_ref.detach().float() * get_fp8_descale_factor(dQ_amax, torch_otype)
dK_ref_float = dK_ref.detach().float() * get_fp8_descale_factor(dK_amax, torch_otype)
dV_ref_float = dV_ref.detach().float() * get_fp8_descale_factor(dV_amax, torch_otype)

if is_ragged:
t_idx_q = sum(seq_len_q_list)
dQ_out[t_idx_q:] = 0
dQ_ref_float[t_idx_q:] = 0
t_idx_kv = sum(seq_len_kv_list)
dK_out[t_idx_kv:] = 0
dK_ref_float[t_idx_kv:] = 0
dV_out[t_idx_kv:] = 0
dV_ref_float[t_idx_kv:] = 0

atol, rtol = 0.04, 0.2
torch.testing.assert_close(dQ_out, dQ_ref_float, atol=atol, rtol=rtol)
torch.testing.assert_close(dK_out, dK_ref_float, atol=atol, rtol=rtol)
torch.testing.assert_close(dV_out, dV_ref_float, atol=atol, rtol=rtol)

if with_sink_token:
torch.testing.assert_close(dSink_token_gpu, dSink_token_ref, atol=0.02, rtol=0.2)
if not perf:
if is_ragged:
dQ_ref = convert_uniform_to_packed(torch.einsum("bshd->bhsd", dQ_ref), seq_len_q_ref, max_t_q)
dK_ref = convert_uniform_to_packed(torch.einsum("bshd->bhsd", dK_ref), seq_len_kv_ref, max_t_kv)
dV_ref = convert_uniform_to_packed(torch.einsum("bshd->bhsd", dV_ref), seq_len_kv_ref, max_t_kv)

dQ_out = dQ_gpu.detach().float() * get_fp8_descale_factor(dQ_amax, torch_otype)
dK_out = dK_gpu.detach().float() * get_fp8_descale_factor(dK_amax, torch_otype)
dV_out = dV_gpu.detach().float() * get_fp8_descale_factor(dV_amax, torch_otype)

dQ_ref_float = dQ_ref.detach().float() * get_fp8_descale_factor(dQ_amax, torch_otype)
dK_ref_float = dK_ref.detach().float() * get_fp8_descale_factor(dK_amax, torch_otype)
dV_ref_float = dV_ref.detach().float() * get_fp8_descale_factor(dV_amax, torch_otype)

if is_ragged:
t_idx_q = sum(seq_len_q_list)
dQ_out[t_idx_q:] = 0
dQ_ref_float[t_idx_q:] = 0
t_idx_kv = sum(seq_len_kv_list)
dK_out[t_idx_kv:] = 0
dK_ref_float[t_idx_kv:] = 0
dV_out[t_idx_kv:] = 0
dV_ref_float[t_idx_kv:] = 0

atol, rtol = 0.04, 0.2
torch.testing.assert_close(dQ_out, dQ_ref_float, atol=atol, rtol=rtol)
torch.testing.assert_close(dK_out, dK_ref_float, atol=atol, rtol=rtol)
torch.testing.assert_close(dV_out, dV_ref_float, atol=atol, rtol=rtol)

if with_sink_token:
torch.testing.assert_close(dSink_token_gpu, dSink_token_ref, atol=0.02, rtol=0.2)

# Print hash and stats for determinism verification
print_tensor_stats(o_gpu, tag="o_gpu")
Expand Down
Loading