rewrite D2 regression test: match existing Stage D1 test pattern with cute.compile + PV tiles
This commit is contained in:
@@ -1,54 +1,166 @@
|
||||
"""Quick test: n_h=1 regression after grid changes."""
|
||||
"""
|
||||
FMHA D2 regression test (matches existing test pattern).
|
||||
|
||||
Uses the same cute.compile + PV tile iteration as test_fmha_v3_stage_d1.py.
|
||||
|
||||
Run: ~/.openclaw/workspace/fire_b200_test tests/unit/test_d2_regression.py
|
||||
"""
|
||||
import torch
|
||||
import math
|
||||
import cutlass
|
||||
import cutlass.cute as cute
|
||||
import cuda.bindings.driver as cuda
|
||||
import cutlass.torch as ct
|
||||
from cutlass import Float32, BFloat16
|
||||
import cuda.bindings.driver as cuda
|
||||
|
||||
from dsv4.kernels.attention.fmha import FmhaKernel
|
||||
|
||||
|
||||
def test():
|
||||
print("=== n_h=1 regression (hd=64, s_k=128) ===")
|
||||
torch.manual_seed(42)
|
||||
M, s_k, hd = 128, 128, 64
|
||||
scale = 1.0 / math.sqrt(hd)
|
||||
|
||||
q = torch.randn(M, hd, 1, dtype=torch.bfloat16, device='cuda')
|
||||
k = torch.randn(s_k, hd, 1, dtype=torch.bfloat16, device='cuda')
|
||||
v = torch.randn(s_k, hd, 1, dtype=torch.bfloat16, device='cuda')
|
||||
o = torch.zeros(M, hd, 1, dtype=torch.bfloat16, device='cuda')
|
||||
|
||||
fmha = FmhaKernel(head_dim=hd, s_k=s_k, normalize=False)
|
||||
stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
|
||||
|
||||
q_c = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
|
||||
k_c = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k))
|
||||
v_c = ct.from_dlpack(v).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v))
|
||||
o_c = ct.from_dlpack(o).mark_layout_dynamic(leading_dim=ct.get_leading_dim(o))
|
||||
lse = torch.zeros(M, dtype=torch.float32, device='cuda')
|
||||
lse_c = ct.from_dlpack(lse).mark_layout_dynamic(leading_dim=ct.get_leading_dim(lse))
|
||||
fmha(q_c, k_c, v_c, o_c, stream, lse_c)
|
||||
|
||||
# External normalization using LSE
|
||||
row_sum = lse.exp()
|
||||
o_norm = o[:,:,0] / row_sum.unsqueeze(-1)
|
||||
|
||||
# Reference
|
||||
scores = torch.matmul(q[:,:,0].float(), k[:,:,0].float().T) * scale
|
||||
def reference_fmha(q, k, v, scale):
|
||||
"""FP32 reference: q (M, hd), k (s_k, hd), v (s_k, hd) → o (M, hd)"""
|
||||
scores = torch.matmul(q.float(), k.float().T) * scale
|
||||
max_s = scores.max(dim=-1, keepdim=True).values
|
||||
exp_s = (scores - max_s).exp()
|
||||
sum_s = exp_s.sum(dim=-1, keepdim=True)
|
||||
p = exp_s / sum_s
|
||||
ref = torch.matmul(p, v[:,:,0].float()).to(torch.bfloat16)
|
||||
o = torch.matmul(p, v.float())
|
||||
return o.to(torch.bfloat16), (sum_s.log() + max_s)
|
||||
|
||||
|
||||
def test_d2_regression():
|
||||
"""Regression test matching existing Stage D1 pattern."""
|
||||
print("\n=== Regression test (hd=64, s_k=128) ===")
|
||||
torch.manual_seed(42)
|
||||
m = 128; n_kv = 128; hd = 64
|
||||
scale = 1.0 / math.sqrt(hd)
|
||||
|
||||
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
|
||||
k = torch.randn(n_kv, hd, 1, dtype=torch.bfloat16, device='cuda')
|
||||
v = torch.randn(n_kv, hd, dtype=torch.bfloat16, device='cuda')
|
||||
|
||||
kernel = FmhaKernel(head_dim=hd, s_k=n_kv, use_smem_p=False)
|
||||
pv_n_tile = kernel.pv_n_tile
|
||||
n_pv_tiles = kernel.n_pv_tiles
|
||||
|
||||
stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
|
||||
|
||||
# Compile with first PV tile
|
||||
v_tile = v[:, 0:pv_n_tile].contiguous()
|
||||
v_kernel = v_tile.unsqueeze(-1)
|
||||
c_tile = torch.zeros(m, pv_n_tile, 1, dtype=torch.bfloat16, device='cuda')
|
||||
lse_tensor = torch.zeros(m, 1, 1, dtype=torch.float32, device='cuda')
|
||||
|
||||
mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
|
||||
mK = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k))
|
||||
mV = ct.from_dlpack(v_kernel).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v_kernel))
|
||||
mC = ct.from_dlpack(c_tile).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c_tile))
|
||||
mLSE = ct.from_dlpack(lse_tensor).mark_layout_dynamic(leading_dim=ct.get_leading_dim(lse_tensor))
|
||||
|
||||
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream, mLSE)
|
||||
|
||||
# Run PV tiles
|
||||
o_unnorm = torch.zeros(m, hd, dtype=torch.float32, device='cuda')
|
||||
for pv in range(n_pv_tiles):
|
||||
v_tile = v[:, pv*pv_n_tile:(pv+1)*pv_n_tile].contiguous()
|
||||
v_kernel = v_tile.unsqueeze(-1)
|
||||
c_tile.zero_()
|
||||
lse_tensor.zero_()
|
||||
|
||||
mV = ct.from_dlpack(v_kernel).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v_kernel))
|
||||
mC = ct.from_dlpack(c_tile).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c_tile))
|
||||
mLSE = ct.from_dlpack(lse_tensor).mark_layout_dynamic(leading_dim=ct.get_leading_dim(lse_tensor))
|
||||
|
||||
compiled(mQ, mK, mV, mC, stream, mLSE)
|
||||
|
||||
o_unnorm[:, pv*pv_n_tile:(pv+1)*pv_n_tile] = c_tile[:,:,0].float()
|
||||
|
||||
# External normalization using LSE
|
||||
lse = lse_tensor[:,0,0] # (m,)
|
||||
row_sum = lse.exp()
|
||||
o_norm = o_unnorm / row_sum.unsqueeze(-1)
|
||||
o_bf16 = o_norm.to(torch.bfloat16)
|
||||
|
||||
# Reference
|
||||
ref, _ = reference_fmha(q[:,:,0], k[:,:,0], v, scale)
|
||||
|
||||
cos = torch.nn.functional.cosine_similarity(
|
||||
o_norm.flatten().float().unsqueeze(0), ref.flatten().float().unsqueeze(0)
|
||||
o_bf16.flatten().float().unsqueeze(0), ref.flatten().float().unsqueeze(0)
|
||||
).item()
|
||||
print(f" cos (ext norm) = {cos:.6f}")
|
||||
if cos >= 0.99:
|
||||
print(" ✅ PASS")
|
||||
else:
|
||||
print(f" ❌ FAIL")
|
||||
print(f" cos = {cos:.6f}")
|
||||
assert cos >= 0.99, f"cosine too low: {cos}"
|
||||
print(" ✅ PASS")
|
||||
|
||||
|
||||
def test_d2_headpacked_128():
|
||||
"""n_h=128, T=1 (Pro decode): M=128, heads packed into M."""
|
||||
print("\n=== n_h=128, T=1 (Pro decode, hd=64) ===")
|
||||
torch.manual_seed(42)
|
||||
n_h, T, s_k, hd = 128, 1, 128, 64
|
||||
scale = 1.0 / math.sqrt(hd)
|
||||
|
||||
# Per-head Q
|
||||
q_heads = torch.randn(n_h, T, hd, dtype=torch.bfloat16, device='cuda')
|
||||
# Pack heads into M: (n_h*T, hd) → (128, 64, 1)
|
||||
q = q_heads.reshape(n_h * T, hd).unsqueeze(-1)
|
||||
k = torch.randn(s_k, hd, 1, dtype=torch.bfloat16, device='cuda')
|
||||
v = torch.randn(s_k, hd, dtype=torch.bfloat16, device='cuda')
|
||||
|
||||
kernel = FmhaKernel(head_dim=hd, s_k=s_k, use_smem_p=False)
|
||||
pv_n_tile = kernel.pv_n_tile
|
||||
n_pv_tiles = kernel.n_pv_tiles
|
||||
|
||||
stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
|
||||
|
||||
v_tile = v[:, 0:pv_n_tile].contiguous().unsqueeze(-1)
|
||||
c_tile = torch.zeros(n_h * T, pv_n_tile, 1, dtype=torch.bfloat16, device='cuda')
|
||||
lse_tensor = torch.zeros(n_h * T, 1, 1, dtype=torch.float32, device='cuda')
|
||||
|
||||
mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
|
||||
mK = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k))
|
||||
mV = ct.from_dlpack(v_tile).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v_tile))
|
||||
mC = ct.from_dlpack(c_tile).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c_tile))
|
||||
mLSE = ct.from_dlpack(lse_tensor).mark_layout_dynamic(leading_dim=ct.get_leading_dim(lse_tensor))
|
||||
|
||||
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream, mLSE)
|
||||
|
||||
o_unnorm = torch.zeros(n_h * T, hd, dtype=torch.float32, device='cuda')
|
||||
for pv in range(n_pv_tiles):
|
||||
v_tile = v[:, pv*pv_n_tile:(pv+1)*pv_n_tile].contiguous().unsqueeze(-1)
|
||||
c_tile.zero_()
|
||||
lse_tensor.zero_()
|
||||
|
||||
mV = ct.from_dlpack(v_tile).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v_tile))
|
||||
mC = ct.from_dlpack(c_tile).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c_tile))
|
||||
mLSE = ct.from_dlpack(lse_tensor).mark_layout_dynamic(leading_dim=ct.get_leading_dim(lse_tensor))
|
||||
|
||||
compiled(mQ, mK, mV, mC, stream, mLSE)
|
||||
|
||||
o_unnorm[:, pv*pv_n_tile:(pv+1)*pv_n_tile] = c_tile[:,:,0].float()
|
||||
|
||||
lse = lse_tensor[:,0,0]
|
||||
row_sum = lse.exp()
|
||||
o_norm = o_unnorm / row_sum.unsqueeze(-1)
|
||||
o_bf16 = o_norm.to(torch.bfloat16)
|
||||
|
||||
# Per-head reference
|
||||
o_ref = torch.zeros(n_h, T, hd, dtype=torch.bfloat16, device='cuda')
|
||||
for h in range(n_h):
|
||||
o_ref[h, 0], _ = reference_fmha(q_heads[h], k[:,:,0], v, scale)
|
||||
o_ref_flat = o_ref.reshape(n_h * T, hd)
|
||||
|
||||
cos = torch.nn.functional.cosine_similarity(
|
||||
o_bf16.flatten().float().unsqueeze(0), o_ref_flat.flatten().float().unsqueeze(0)
|
||||
).item()
|
||||
print(f" cos = {cos:.6f}")
|
||||
assert cos >= 0.99, f"cosine too low: {cos}"
|
||||
print(" ✅ PASS")
|
||||
|
||||
|
||||
def test():
|
||||
print("=== D2: Head-Packed FMHA ===")
|
||||
test_d2_regression()
|
||||
test_d2_headpacked_128()
|
||||
print("\n=== ALL TESTS PASSED ===")
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
|
||||
Reference in New Issue
Block a user