253 lines
10 KiB
Python
253 lines
10 KiB
Python
#!/usr/bin/env python3
|
||
"""
|
||
Test NVFP4 attention: quantize Q and K, GEMM in NVFP4, softmax in BF16.
|
||
|
||
Step 1: Verify NVFP4 quantize/dequant roundtrip for attention
|
||
Step 2: Q×K^T using CuTeDSL NVFP4 GEMM
|
||
Step 3: Softmax + attn×V
|
||
Step 4: Full pipeline with real weights, compare to BF16 SDPA
|
||
|
||
Usage (on B200):
|
||
cd /root/nvfp4-megamoe-kernel
|
||
PYTHONPATH=/root/nvfp4-megamoe-kernel tests/venv/bin/python tests/test_nvfp4_attention_b200.py
|
||
"""
|
||
|
||
import sys, os, json, torch, torch.nn.functional as F, math
|
||
from safetensors import safe_open
|
||
|
||
REPO = "/root/nvfp4-megamoe-kernel"
|
||
sys.path.insert(0, REPO)
|
||
MODEL = "/root/nvidia-meeting/DeepSeek-V4-Pro-NVFP4"
|
||
DEV = "cuda:0"
|
||
|
||
H = 7168; NH = 128; HD = 512; NOPE = 448; ROPE = 64
|
||
QL = 1536; OL = 1024; OG = 16; HPG = NH // OG
|
||
EPS = 1e-6; WINDOW = 8192; SCALE = HD ** -0.5
|
||
|
||
E2M1 = torch.tensor([0,.5,1.,1.5,2.,3.,4.,6.,-0,-.5,-1.,-1.5,-2.,-3.,-4.,-6.], dtype=torch.float32)
|
||
|
||
_cache = {}
|
||
def P(k, wm, md):
|
||
if k in _cache: return _cache[k]
|
||
with safe_open(os.path.join(md, wm[k]), framework="pt") as f:
|
||
t = f.get_tensor(k)
|
||
_cache[k] = t
|
||
return t
|
||
|
||
def dequant(w, sf, gs):
|
||
d = w.device; lut = E2M1.to(d)
|
||
lo = lut[(w & 0xF).long()]; hi = lut[((w >> 4) & 0xF).long()]
|
||
O, I2 = w.shape; I = I2*2
|
||
u = torch.empty(O, I, dtype=torch.float32, device=d)
|
||
u[:,0::2] = lo; u[:,1::2] = hi
|
||
bs = sf.float().repeat_interleave(16, dim=1)[:O,:I]
|
||
return (u * bs * gs).to(torch.bfloat16)
|
||
|
||
def rms(x, w, eps=1e-6):
|
||
v = x.float().pow(2).mean(-1, keepdim=True)
|
||
return (w.float() * (x * torch.rsqrt(v+eps)).float()).to(x.dtype)
|
||
|
||
def make_runner(w, sf, gs_t, inf, outf, fused=False, lw=None):
|
||
from cutedsl.nvfp4_linear import CuTeDSLNvfp4Linear
|
||
fp4 = w.view(torch.float4_e2m1fn_x2).permute(1,0).contiguous()
|
||
s = sf.to(torch.float8_e4m3fn) if sf.dtype != torch.float8_e4m3fn else sf
|
||
s = s.permute(1,0).contiguous()
|
||
if fused and gs_t.numel() == 2:
|
||
g1,g2 = gs_t[0].item(), gs_t[1].item(); gs = max(g1,g2)
|
||
if g1 != g2:
|
||
s32 = s.float(); sp = lw[0] if lw else outf//2
|
||
s32[:sp] *= g1/gs; s32[sp:] *= g2/gs; s = s32.to(torch.float8_e4m3fn)
|
||
else:
|
||
gs = gs_t.max().item() if gs_t.numel() > 1 else gs_t.item()
|
||
r = CuTeDSLNvfp4Linear(in_features=inf, out_features=outf, max_num_tokens=8192, device=str(w.device))
|
||
r.fp4 = [fp4]; r.sf = [s]; r.gs = [gs]
|
||
r.finalize_weights(); r._ensure_initialized()
|
||
return r
|
||
|
||
def apply_gptj_rope(x, positions, cos_sin, nope, rope):
|
||
if rope == 0 or x.numel() == 0: return x
|
||
half = rope // 2
|
||
cos = cos_sin[positions, :half].to(x.dtype)
|
||
sin = cos_sin[positions, half:].to(x.dtype)
|
||
if x.dim() == 3: cos = cos.unsqueeze(1); sin = sin.unsqueeze(1)
|
||
x_rope = x[..., nope:].clone()
|
||
even = x_rope[..., 0::2]; odd = x_rope[..., 1::2]
|
||
out = x.clone()
|
||
out[..., nope:][..., 0::2] = even * cos - odd * sin
|
||
out[..., nope:][..., 1::2] = even * sin + odd * cos
|
||
return out
|
||
|
||
def build_cos_sin(max_pos=4096, rope_dim=ROPE):
|
||
half = rope_dim // 2
|
||
inv_freq = 1.0 / (10000.0 ** (torch.arange(0, half, dtype=torch.float32) / half))
|
||
freqs = torch.outer(torch.arange(max_pos, dtype=torch.float32), inv_freq)
|
||
return torch.cat([freqs.cos(), freqs.sin()], dim=-1)
|
||
|
||
|
||
def bf16_full_attention(q, kv, scale):
|
||
"""BF16 reference: full self-attention with causal mask."""
|
||
T, NH, HD = q.shape
|
||
q_2d = q.reshape(T * NH, HD)
|
||
kv_expanded = kv.unsqueeze(1).expand(-1, NH, -1).contiguous()
|
||
k_2d = kv_expanded.permute(1, 0, 2).unsqueeze(1).expand(NH, T, T, -1).contiguous().reshape(T * NH, T, HD)
|
||
v_2d = k_2d.clone()
|
||
scores = torch.matmul(q_2d.unsqueeze(1), k_2d.transpose(-1, -2)) * scale
|
||
query_pos = torch.arange(T, device=q.device).unsqueeze(1).repeat(1, NH).reshape(T * NH)
|
||
kv_pos = torch.arange(T, device=q.device).unsqueeze(0)
|
||
causal = kv_pos <= query_pos.unsqueeze(1)
|
||
scores = scores.squeeze(1).masked_fill(~causal, float('-inf'))
|
||
weights = F.softmax(scores.float(), dim=-1).to(q.dtype)
|
||
out = torch.matmul(weights.unsqueeze(1), v_2d).squeeze(1)
|
||
return out.reshape(T, NH, HD)
|
||
|
||
|
||
def nvfp4_qk_attention(q, kv, scale):
|
||
"""NVFP4 attention: quantize Q and K for Q×K^T, then BF16 softmax + attn×V.
|
||
|
||
Key insight: Q×K^T is (T*NH, HD) × (HD, T) = (T*NH, T).
|
||
This is a standard GEMM that CuTeDSL can handle.
|
||
We quantize Q as the "activation" and K^T as the "weight".
|
||
"""
|
||
from cutedsl.bridge import quantize_to_nvfp4, quantize_activation_nvfp4
|
||
from cutedsl.nvfp4_linear import CuTeDSLNvfp4Linear
|
||
|
||
T, NH, HD = q.shape
|
||
device = q.device
|
||
|
||
# Q as activation: (T*NH, HD) → NVFP4
|
||
q_2d = q.reshape(T * NH, HD)
|
||
q_fp4, q_sf, q_gs = quantize_to_nvfp4(q_2d) # (T*NH, HD//2), (T*NH, HD//16), scalar
|
||
|
||
# K as weight: (T, HD) → transpose to (HD, T), quantize as weight
|
||
# In our framework, "weight" means quantized along K dim
|
||
kv_T = kv.T.contiguous() # (HD, T)
|
||
w_fp4, w_sf, w_gs = quantize_to_nvfp4(kv_T) # (HD//2, T), (HD//16, T), scalar
|
||
|
||
# Use CuTeDSLNvfp4Linear runner for Q×K^T GEMM
|
||
# in_features=HD, out_features=T
|
||
# Q is "activation" side, K^T is "weight" side
|
||
M = T * NH
|
||
K = HD
|
||
N = T
|
||
|
||
# Create runner for this specific (M, K, N) combination
|
||
runner = CuTeDSLNvfp4Linear(
|
||
in_features=K, out_features=N, max_num_tokens=M, device=str(device)
|
||
)
|
||
|
||
# Weight is kv_T: set up as (N, K//2) in N-major (standard row-major)
|
||
# runner expects: weight fp4 is (N, K//2), weight sf is (N, K//16)
|
||
# Our w_fp4 from quantize_to_nvfp4(kv_T) is (K//2, T) — that's (K_packed, N)
|
||
# Need to transpose to (N, K_packed)
|
||
w_fp4_loaded = w_fp4.T.contiguous() # (T, HD//2) = (N, K_packed)
|
||
w_sf_loaded = w_sf.T.contiguous() # (T, HD//16) = (N, K_sf)
|
||
|
||
runner.fp4 = [w_fp4_loaded]
|
||
runner.sf = [w_sf_loaded]
|
||
runner.gs = [w_gs]
|
||
runner.finalize_weights()
|
||
runner._ensure_initialized()
|
||
|
||
# Run: Q×K^T
|
||
# q_2d is (M, K) BF16, runner produces (M, N) BF16
|
||
scores = runner.run(q_2d) * scale # (T*NH, T)
|
||
|
||
# Causal mask
|
||
query_pos = torch.arange(T, device=device).unsqueeze(1).repeat(1, NH).reshape(T * NH)
|
||
kv_pos = torch.arange(T, device=device).unsqueeze(0)
|
||
causal = kv_pos <= query_pos.unsqueeze(1)
|
||
scores = scores.masked_fill(~causal, float('-inf'))
|
||
|
||
# Softmax in BF16 (must be full precision for numerical stability)
|
||
weights = F.softmax(scores.float(), dim=-1).to(q.dtype) # (T*NH, T)
|
||
|
||
# attn×V: (T*NH, T) × (T, HD) → (T*NH, HD)
|
||
# V = kv (shared, BF16) — no quantization needed here since attn weights are already BF16
|
||
out = torch.matmul(weights, kv) # (T*NH, HD)
|
||
|
||
return out.reshape(T, NH, HD)
|
||
|
||
|
||
def main():
|
||
torch.cuda.set_device(0)
|
||
torch.manual_seed(42)
|
||
|
||
print("=" * 70)
|
||
print(" NVFP4 Attention Kernel Test")
|
||
print("=" * 70)
|
||
|
||
with open(os.path.join(MODEL, "model.safetensors.index.json")) as f:
|
||
wm = json.load(f)["weight_map"]
|
||
G = lambda k: P(k, wm, MODEL).to(DEV)
|
||
|
||
p = "model.layers.0"; a = f"{p}.self_attn"
|
||
|
||
# Load weights
|
||
emb = G("model.embed_tokens.weight")
|
||
anorm = G(f"{p}.input_layernorm.weight")
|
||
qn = G(f"{a}.q_a_norm.weight"); kvn = G(f"{a}.kv_norm.weight")
|
||
woa = G(f"{a}.o_a_proj.weight")
|
||
|
||
qa_w = G(f"{a}.q_a_proj.weight"); qa_sf = G(f"{a}.q_a_proj.weight_scale"); qa_gs = G(f"{a}.q_a_proj.weight_scale_2")
|
||
qb_w = G(f"{a}.q_b_proj.weight"); qb_sf = G(f"{a}.q_b_proj.weight_scale"); qb_gs = G(f"{a}.q_b_proj.weight_scale_2")
|
||
kv_w = G(f"{a}.kv_proj.weight"); kv_sf = G(f"{a}.kv_proj.weight_scale"); kv_gs = G(f"{a}.kv_proj.weight_scale_2")
|
||
wob_w = G(f"{a}.o_b_proj.weight"); wob_sf = G(f"{a}.o_b_proj.weight_scale"); wob_gs = G(f"{a}.o_b_proj.weight_scale_2")
|
||
sinks = G(f"{a}.sinks")
|
||
|
||
# BF16 references
|
||
qa_bf16 = dequant(qa_w, qa_sf, qa_gs.item())
|
||
qb_bf16 = dequant(qb_w, qb_sf, qb_gs.item())
|
||
kv_bf16 = dequant(kv_w, kv_sf, kv_gs.item())
|
||
wob_bf16 = dequant(wob_w, wob_sf, wob_gs.item())
|
||
|
||
# CuTeDSL runners
|
||
r_qa = make_runner(qa_w, qa_sf, qa_gs, H, qa_w.shape[0])
|
||
r_qb = make_runner(qb_w, qb_sf, qb_gs, QL, qb_w.shape[0])
|
||
r_kv = make_runner(kv_w, kv_sf, kv_gs, H, kv_w.shape[0])
|
||
r_wob = make_runner(wob_w, wob_sf, wob_gs, OG*OL, wob_w.shape[0])
|
||
|
||
# Input
|
||
token_ids = torch.tensor([1, 450, 8403, 315, 5413, 374], dtype=torch.long, device=DEV)
|
||
NT = len(token_ids)
|
||
cos_sin = build_cos_sin(max_pos=WINDOW + 256).to(DEV)
|
||
positions = torch.arange(NT, dtype=torch.int64, device=DEV)
|
||
|
||
print(f" Input: {NT} tokens, {NH} heads, HD={HD}")
|
||
|
||
with torch.no_grad():
|
||
hidden = emb[token_ids]
|
||
normed = rms(hidden, anorm, EPS)
|
||
|
||
# Projections
|
||
qa_cute = r_qa.run(normed)
|
||
kv_cute = r_kv.run(normed)
|
||
qa_n = rms(qa_cute, qn, EPS)
|
||
kv_n = rms(kv_cute, kvn, EPS)
|
||
q_cute = r_qb.run(qa_n).view(NT, NH, HD)
|
||
q_rope = apply_gptj_rope(q_cute, positions, cos_sin, NOPE, ROPE)
|
||
|
||
# ── BF16 reference ────────────────────────────────────────────
|
||
print("\n--- Step 1: BF16 reference attention ---")
|
||
o_bf16 = bf16_full_attention(q_rope, kv_n, SCALE)
|
||
print(f" BF16 attention output: amax={o_bf16.amax():.4f} NaN={torch.isnan(o_bf16).any()}")
|
||
|
||
# ── NVFP4 Q×K^T attention ────────────────────────────────────
|
||
print("\n--- Step 2: NVFP4 Q×K^T attention ---")
|
||
try:
|
||
o_nvfp4 = nvfp4_qk_attention(q_rope, kv_n, SCALE)
|
||
print(f" NVFP4 attention output: amax={o_nvfp4.amax():.4f} NaN={torch.isnan(o_nvfp4).any()}")
|
||
|
||
c = F.cosine_similarity(o_nvfp4.flatten().unsqueeze(0).float(), o_bf16.flatten().unsqueeze(0).float()).item()
|
||
print(f" NVFP4 vs BF16 cosine: {c:.6f} {'✅' if c>=0.98 else '❌'}")
|
||
except Exception as e:
|
||
print(f" ERROR: {e}")
|
||
import traceback; traceback.print_exc()
|
||
|
||
print("\n" + "=" * 70)
|
||
print(" Done")
|
||
print("=" * 70)
|
||
|
||
|
||
if __name__ == "__main__":
|
||
main()
|