Files
nvfp4-megamoe-kernel/tests/test_kv_cache_b200.py

357 lines
15 KiB
Python

#!/usr/bin/env python3
"""
DeepSeek-V4 KV Cache Kernel — NVFP4 Compressed Storage
Architecture:
- SWA cache: (T, HD=512) per token, stored as fp8_e4m3 (512 bytes per token)
- CSA cache (C4A): every 4th token stored, (T//4, HD) fp8 (128 bytes per token)
- HCA cache (C128A): every 128th token stored, (T//128, HD) fp8 (4 bytes per token)
The KV latent is (1, HD=512) — single KV head. After kv_norm + RoPE,
it's quantized to fp8_e4m3 and stored in the paged KV cache.
For CSA/HCA layers, the compressor further reduces the cache:
- The indexer finds top-k positions in the compressed cache
- Attention only attends to those positions
This kernel tests:
1. KV quantization: BF16 → fp8_e4m3 (with per-token scale)
2. KV dequantization: fp8_e4m3 → BF16
3. RoPE on dequantized KV (applied after dequant)
4. Full attention using the cache
5. Compressed cache (CSA/HCA) storage and retrieval
Usage (on B200):
cd /root/nvfp4-megamoe-kernel
PYTHONPATH=/root/nvfp4-megamoe-kernel tests/venv/bin/python tests/test_kv_cache_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 = 128; 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 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 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:2*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
# ── KV Cache Kernels ────────────────────────────────────────────────
def kv_quantize_fp8(kv_bf16):
"""Quantize KV latent to fp8_e4m3 with per-token scale.
kv_bf16: (T, HD) BF16
Returns: (T, HD) fp8, (T, 1) per-token scale (BF16)
"""
# Per-token absmax
amax = kv_bf16.float().abs().amax(dim=-1, keepdim=True).clamp(min=1e-12)
fp8_max = torch.tensor(448.0, dtype=torch.float32, device=kv_bf16.device) # e4m3 max
scale = fp8_max / amax # (T, 1)
kv_scaled = kv_bf16.float() * scale
kv_fp8 = kv_scaled.to(torch.float8_e4m3fn)
# Store inverse scale for dequant
inv_scale = amax / fp8_max # (T, 1) — multiply by this to recover
return kv_fp8, inv_scale.to(torch.bfloat16)
def kv_dequantize_fp8(kv_fp8, inv_scale):
"""Dequantize fp8 KV back to BF16.
kv_fp8: (T, HD) fp8_e4m3
inv_scale: (T, 1) per-token scale
Returns: (T, HD) BF16
"""
return (kv_fp8.to(torch.bfloat16) * inv_scale).to(torch.bfloat16)
def kv_quantize_nvfp4(kv_bf16):
"""Quantize KV latent to NVFP4 using CuTeDSL quantize_to_nvfp4.
More aggressive compression: 2x smaller than fp8 (4 bits vs 8 bits per element).
kv_bf16: (T, HD) BF16
Returns: (T, HD//2) fp4, (T, HD//16) sf, scalar gs
"""
from cutedsl.bridge import quantize_to_nvfp4
return quantize_to_nvfp4(kv_bf16)
def kv_dequantize_nvfp4(kv_fp4, kv_sf, kv_gs, head_dim=HD):
"""Dequantize NVFP4 KV back to BF16.
kv_fp4: (T, HD//2) fp4 (as float4_e2m1fn_x2 viewed as uint8)
kv_sf: (T, HD//16) fp8 block scales
kv_gs: scalar global scale
"""
device = kv_fp4.device
lut = E2M1.to(device)
packed = kv_fp4.view(torch.uint8)
lo = lut[(packed & 0xF).long()]
hi = lut[((packed >> 4) & 0xF).long()]
T = kv_fp4.shape[0]
u = torch.empty(T, head_dim, dtype=torch.float32, device=device)
u[:, 0::2] = lo
u[:, 1::2] = hi
sf_exp = kv_sf.float().repeat_interleave(16, dim=1)[:, :head_dim]
return (u * sf_exp * kv_gs).to(torch.bfloat16)
def paged_kv_write(kv_fp8, slot_mapping, cache, block_size):
"""Write KV into paged cache.
kv_fp8: (T, HD) fp8 to write
slot_mapping: (T,) slot indices (position in flat cache)
cache: (num_blocks, block_size, HD) fp8 cache tensor
block_size: tokens per block
"""
for t in range(kv_fp8.shape[0]):
slot = slot_mapping[t].item()
block_idx = slot // block_size
offset = slot % block_size
if block_idx < cache.shape[0]:
cache[block_idx, offset] = kv_fp8[t]
def paged_kv_read(slot_mapping, cache, block_size, num_tokens):
"""Read KV from paged cache.
Returns: (num_tokens, HD) fp8
"""
device = cache.device
HD = cache.shape[-1]
kv = torch.zeros(num_tokens, HD, dtype=cache.dtype, device=device)
for t in range(num_tokens):
slot = slot_mapping[t].item()
block_idx = slot // block_size
offset = slot % block_size
if block_idx < cache.shape[0]:
kv[t] = cache[block_idx, offset]
return kv
def main():
torch.cuda.set_device(0)
torch.manual_seed(42)
print("=" * 70)
print(" DeepSeek-V4 KV Cache Kernel Test")
print(" fp8 and NVFP4 quantization for paged KV cache")
print("=" * 70)
# Load real weights
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"
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")
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")
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")
r_qa = make_runner(qa_w, qa_sf, qa_gs, H, qa_w.shape[0])
r_kv = make_runner(kv_w, kv_sf, kv_gs, H, kv_w.shape[0])
cos_sin = build_cos_sin(max_pos=4096).to(DEV)
token_ids = torch.tensor([1, 450, 8403, 315, 5413, 374], dtype=torch.long, device=DEV)
NT = len(token_ids)
positions = torch.arange(NT, dtype=torch.int64, device=DEV)
with torch.no_grad():
hidden = emb[token_ids]
normed = rms(hidden, anorm, EPS)
kv_bf16 = r_kv.run(normed)
kv_bf16 = rms(kv_bf16, kvn, EPS)
# ── Test 1: FP8 KV quantize/dequant roundtrip ────────────────
print("\n--- Test 1: FP8 KV quantize/dequant ---")
kv_fp8, inv_scale = kv_quantize_fp8(kv_bf16)
kv_recovered = kv_dequantize_fp8(kv_fp8, inv_scale)
c = F.cosine_similarity(kv_bf16.flatten().unsqueeze(0).float(), kv_recovered.flatten().unsqueeze(0).float()).item()
print(f" FP8 roundtrip cosine: {c:.6f} {'' if c>=0.99 else ''}")
print(f" FP8 cache size: {kv_fp8.numel()} bytes (vs {kv_bf16.numel()*2} BF16)")
# ── Test 2: NVFP4 KV quantize/dequant roundtrip ──────────────
print("\n--- Test 2: NVFP4 KV quantize/dequant ---")
try:
kv_nfp4, kv_nsf, kv_ngs = kv_quantize_nvfp4(kv_bf16)
kv_n_recovered = kv_dequantize_nvfp4(kv_nfp4, kv_nsf, kv_ngs)
c = F.cosine_similarity(kv_bf16.flatten().unsqueeze(0).float(), kv_n_recovered.flatten().unsqueeze(0).float()).item()
print(f" NVFP4 roundtrip cosine: {c:.6f} {'' if c>=0.98 else ''}")
print(f" NVFP4 cache size: {kv_nfp4.view(torch.uint8).numel()} bytes (vs {kv_bf16.numel()*2} BF16, {kv_fp8.numel()} FP8)")
except Exception as e:
print(f" NVFP4 quantize failed: {e}")
# ── Test 3: Paged KV cache write/read with FP8 ───────────────
print("\n--- Test 3: Paged KV cache (FP8) ---")
block_size = 256
num_blocks = 64
cache = torch.zeros(num_blocks, block_size, HD, dtype=torch.float8_e4m3fn, device=DEV)
# Slot mapping: position → flat slot (simplified: slot = position)
slot_mapping = positions # (NT,)
# Write KV into cache
paged_kv_write(kv_fp8, slot_mapping, cache, block_size)
# Read back
kv_read = paged_kv_read(slot_mapping, cache, block_size, NT)
c = F.cosine_similarity(kv_fp8.flatten().unsqueeze(0).float(), kv_read.flatten().unsqueeze(0).float()).item()
print(f" Paged read back cosine: {c:.6f} {'' if c>=0.999 else ''}")
# ── Test 4: Apply RoPE after dequant ─────────────────────────
print("\n--- Test 4: RoPE on dequantized KV ---")
# KV needs RoPE applied at the positions it was stored at
kv_with_rope = apply_gptj_rope(kv_recovered.unsqueeze(1), positions, cos_sin, NOPE, ROPE).squeeze(1)
print(f" KV+RoPE: amax={kv_with_rope.amax():.4f} NaN={torch.isnan(kv_with_rope).any()}")
# ── Test 5: Full attention with FP8 KV cache ─────────────────
print("\n--- Test 5: Full attention pipeline with FP8 KV cache ---")
qa_bf16_ref = dequant(qa_w, qa_sf, qa_gs.item())
qb_bf16_ref = dequant(
G(f"{a}.q_b_proj.weight"),
G(f"{a}.q_b_proj.weight_scale"),
G(f"{a}.q_b_proj.weight_scale_2").item()
)
kv_bf16_ref = dequant(kv_w, kv_sf, kv_gs.item())
r_qb = make_runner(
G(f"{a}.q_b_proj.weight"),
G(f"{a}.q_b_proj.weight_scale"),
G(f"{a}.q_b_proj.weight_scale_2"),
QL, G(f"{a}.q_b_proj.weight").shape[0]
)
# Full BF16 reference
qa_ref = normed @ qa_bf16_ref.T
kv_ref = normed @ kv_bf16_ref.T
qa_n_ref = rms(qa_ref, qn, EPS)
kv_n_ref = rms(kv_ref, kvn, EPS)
q_ref = (qa_n_ref @ qb_bf16_ref.T).view(NT, NH, HD)
q_rope_ref = apply_gptj_rope(q_ref, positions, cos_sin, NOPE, ROPE)
# BF16 causal attention using dequantized FP8 KV cache
kv_from_cache = kv_dequantize_fp8(kv_read, inv_scale)
kv_from_cache_rope = apply_gptj_rope(kv_from_cache.unsqueeze(1), positions, cos_sin, NOPE, ROPE).squeeze(1)
# Full attention with cached KV
T, NH_t, HD_t = q_rope_ref.shape
q_2d = q_rope_ref.reshape(T * NH_t, HD_t)
kv_exp = kv_from_cache_rope.unsqueeze(1).expand(-1, NH_t, -1).contiguous()
k_2d = kv_exp.permute(1, 0, 2).unsqueeze(1).expand(NH_t, T, T, -1).contiguous().reshape(T * NH_t, T, HD_t)
scores = torch.matmul(q_2d.unsqueeze(1), k_2d.transpose(-1, -2)) * SCALE
qpos = torch.arange(T, device=DEV).unsqueeze(1).repeat(1, NH_t).reshape(T * NH_t)
kpos = torch.arange(T, device=DEV).unsqueeze(0)
causal = kpos <= qpos.unsqueeze(1)
scores = scores.squeeze(1).masked_fill(~causal, float('-inf'))
weights = F.softmax(scores.float(), dim=-1).to(q_rope_ref.dtype)
v_2d = k_2d.clone()
out = torch.matmul(weights.unsqueeze(1), v_2d).squeeze(1).reshape(T, NH_t, HD_t)
# BF16 attention with original (no cache) KV
kv_exp2 = kv_n_ref.unsqueeze(1).expand(-1, NH_t, -1).contiguous()
k_2d2 = kv_exp2.permute(1, 0, 2).unsqueeze(1).expand(NH_t, T, T, -1).contiguous().reshape(T * NH_t, T, HD_t)
scores2 = torch.matmul(q_2d.unsqueeze(1), k_2d2.transpose(-1, -2)) * SCALE
scores2 = scores2.squeeze(1).masked_fill(~causal, float('-inf'))
weights2 = F.softmax(scores2.float(), dim=-1).to(q_rope_ref.dtype)
out2 = torch.matmul(weights2.unsqueeze(1), v_2d).squeeze(1).reshape(T, NH_t, HD_t)
c = F.cosine_similarity(out.flatten().unsqueeze(0).float(), out2.flatten().unsqueeze(0).float()).item()
print(f" FP8 cached KV vs BF16 KV attention cosine: {c:.6f} {'' if c>=0.98 else ''}")
# ── Test 6: CSA compressed cache (cr=4) ──────────────────────
print("\n--- Test 6: CSA compressed cache (cr=4) ---")
cr = 4
# Store every 4th token in the compressed cache
compressed_positions = positions[::cr] # every 4th position
compressed_kv = kv_fp8[::cr] # (T//4, HD) fp8
compressed_inv_scale = inv_scale[::cr]
print(f" Compressed KV shape: {compressed_kv.shape} (from {kv_fp8.shape})")
print(f" Compression ratio: {kv_fp8.shape[0] / compressed_kv.shape[0]:.0f}x")
# Dequant compressed KV
compressed_kv_bf16 = kv_dequantize_fp8(compressed_kv, compressed_inv_scale)
c = F.cosine_similarity(kv_bf16[::cr].flatten().unsqueeze(0).float(), compressed_kv_bf16.flatten().unsqueeze(0).float()).item()
print(f" Compressed KV dequant cosine: {c:.6f} {'' if c>=0.99 else ''}")
# ── Test 7: HCA compressed cache (cr=128) ────────────────────
print("\n--- Test 7: HCA compressed cache (cr=128) ---")
cr = 128
compressed_positions_128 = positions[::cr]
compressed_kv_128 = kv_fp8[::cr] if len(kv_fp8) >= cr else kv_fp8[:1]
compressed_inv_scale_128 = inv_scale[::cr] if len(inv_scale) >= cr else inv_scale[:1]
print(f" HCA compressed KV shape: {compressed_kv_128.shape}")
print(f" Tokens in HCA cache: {compressed_kv_128.shape[0]} (from {NT})")
print(f"\n{'='*70}")
print(f" DONE — KV cache kernels tested")
print(f"{'='*70}")
if __name__ == "__main__":
main()