diff --git a/tests/test_decode_vs_prefill_b200.py b/tests/test_decode_vs_prefill_b200.py new file mode 100644 index 00000000..2d236d33 --- /dev/null +++ b/tests/test_decode_vs_prefill_b200.py @@ -0,0 +1,274 @@ +#!/usr/bin/env python3 +""" +DeepSeek-V4 Decode vs Prefill Consistency Test + +Verifies that: +1. Decode attention (using KV cache) produces the same output as + prefill attention (raw KV) for the same token position +2. The cosine similarity between decode and prefill outputs is > 0.98 + +This is the CRITICAL test: if it passes, the KV cache pipeline is correct +and the vLLM container should produce valid output. + +Usage (on B200): + cd /root/nvfp4-megamoe-kernel + PYTHONPATH=/root/nvfp4-megamoe-kernel tests/venv/bin/python tests/test_decode_vs_prefill_b200.py +""" + +import sys, os, json, torch, torch.nn.functional as F, time +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 +NUM_LAYERS = 61 + +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 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_dim, rope_dim): + if rope_dim == 0 or x.numel() == 0: return x + half = rope_dim // 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_dim:].clone() + even = x_rope[..., 0::2]; odd = x_rope[..., 1::2] + out = x.clone() + out[..., nope_dim:][..., 0::2] = even * cos - odd * sin + out[..., nope_dim:][..., 1::2] = even * sin + odd * cos + return out + +def apply_inv_gptj_rope(x, positions, cos_sin, nope_dim, rope_dim): + if rope_dim == 0 or x.numel() == 0: return x + half = rope_dim // 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_dim:].clone() + even = x_rope[..., 0::2]; odd = x_rope[..., 1::2] + out = x.clone() + out[..., nope_dim:][..., 0::2] = even * cos + odd * sin + out[..., nope_dim:][..., 1::2] = -even * sin + odd * cos + return out + +def kv_quantize_fp8(kv_bf16): + 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) + scale = fp8_max / amax + kv_fp8 = (kv_bf16.float() * scale).to(torch.float8_e4m3fn) + inv_scale = (amax / fp8_max).to(torch.bfloat16) + return kv_fp8, inv_scale + +def kv_dequantize_fp8(kv_fp8, inv_scale): + return (kv_fp8.to(torch.bfloat16) * inv_scale).to(torch.bfloat16) + +def causal_prefill_attention(q, kv, scale): + T, NH, HD = q.shape + q_t = q.permute(1, 0, 2) + kv_exp = kv.unsqueeze(0).expand(NH, -1, -1) + out = F.scaled_dot_product_attention(q_t, kv_exp, kv_exp, is_causal=True, scale=scale) + return out.permute(1, 0, 2) + +def decode_attention(q, kv, scale): + NH = q.shape[1]; HD = q.shape[2] + q_t = q.permute(1, 0, 2) + kv_exp = kv.unsqueeze(0).expand(NH, -1, -1) + out = F.scaled_dot_product_attention(q_t, kv_exp, kv_exp, is_causal=False, scale=scale) + return out.permute(1, 0, 2) + + +def test_layer_decode_vs_prefill(layer_id): + """For a single layer, verify decode matches prefill.""" + torch.cuda.set_device(0) + torch.cuda.empty_cache() + + 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 = f"model.layers.{layer_id}"; a = f"{p}.self_attn" + cr = 128 if layer_id == 0 else (0 if layer_id == 60 else 4) + lt = f"C{cr}A" if cr > 1 else "SWA" + + 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") + + 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]) + cos_sin = build_cos_sin(max_pos=4096).to(DEV) + + # Paged KV cache + block_size = 64; max_tokens = 32; num_blocks = (max_tokens + block_size - 1) // block_size + kv_cache = torch.zeros(num_blocks, block_size, HD, dtype=torch.float8_e4m3fn, device=DEV) + inv_scale_cache = torch.zeros(max_tokens, 1, dtype=torch.bfloat16, device=DEV) + + N = 8 # Prefill tokens + token_ids = torch.tensor([1, 450, 8403, 315, 5413, 374, 2198, 643], dtype=torch.long, device=DEV) + + with torch.no_grad(): + # ── PREFILL: process all N tokens at once ─────────────── + positions_p = torch.arange(N, dtype=torch.int64, device=DEV) + hidden_p = emb[token_ids] + normed_p = rms(hidden_p, anorm, EPS) + qa_p = r_qa.run(normed_p); kv_p = r_kv.run(normed_p) + qa_n_p = rms(qa_p, qn, EPS); kv_n_p = rms(kv_p, kvn, EPS) + q_p = r_qb.run(qa_n_p).view(N, NH, HD) + q_rope_p = apply_gptj_rope(q_p, positions_p, cos_sin, NOPE, ROPE) + kv_rope_p = apply_gptj_rope(kv_n_p.unsqueeze(1), positions_p, cos_sin, NOPE, ROPE).squeeze(1) + + # Write prefill KV to cache + kv_fp8_p, inv_s_p = kv_quantize_fp8(kv_rope_p) + slots_p = positions_p + bi_p = slots_p // block_size; oi_p = slots_p % block_size + kv_cache[bi_p, oi_p] = kv_fp8_p + for t in range(N): + inv_scale_cache[slots_p[t]] = inv_s_p[t] + + # Prefill attention (raw KV) + o_prefill = causal_prefill_attention(q_rope_p, kv_rope_p, SCALE) + o_inv_p = apply_inv_gptj_rope(o_prefill, positions_p, cos_sin, NOPE, ROPE) + o_grp_p = o_inv_p.reshape(N, OG, HPG * HD).permute(1, 0, 2) + woa_3d = woa.view(OG, OL, HPG * HD) + z_p = torch.bmm(o_grp_p, woa_3d.transpose(1, 2)).permute(1, 0, 2).reshape(N, OG * OL) + attn_prefill = r_wob.run(z_p) + + # ── DECODE: process token N (one at a time) ──────────── + decode_id = torch.tensor([991], dtype=torch.long, device=DEV) + pos_d = torch.tensor([N], dtype=torch.int64, device=DEV) + hidden_d = emb[decode_id] + normed_d = rms(hidden_d, anorm, EPS) + qa_d = r_qa.run(normed_d); kv_d = r_kv.run(normed_d) + qa_n_d = rms(qa_d, qn, EPS); kv_n_d = rms(kv_d, kvn, EPS) + q_d = r_qb.run(qa_n_d).view(1, NH, HD) + q_rope_d = apply_gptj_rope(q_d, pos_d, cos_sin, NOPE, ROPE) + kv_rope_d = apply_gptj_rope(kv_n_d.unsqueeze(1), pos_d, cos_sin, NOPE, ROPE).squeeze(1) + + # Write decode KV to cache + kv_fp8_d, inv_s_d = kv_quantize_fp8(kv_rope_d) + slot_d = pos_d[0].item() + bi_d = slot_d // block_size; oi_d = slot_d % block_size + kv_cache[bi_d, oi_d] = kv_fp8_d[0] + inv_scale_cache[slot_d] = inv_s_d[0] + + # Decode attention: read from cache + all_slots = torch.arange(N + 1, dtype=torch.int64, device=DEV) + all_bi = all_slots // block_size; all_oi = all_slots % block_size + kv_cached_fp8 = kv_cache[all_bi, all_oi] + kv_cached_inv = inv_scale_cache[all_slots] + kv_cached = kv_dequantize_fp8(kv_cached_fp8, kv_cached_inv) + + # SWA window + ws = max(0, N - WINDOW + 1) + kv_window = kv_cached[ws:] + o_decode = decode_attention(q_rope_d, kv_window, SCALE) + + # Full output pipeline for decode + o_inv_d = apply_inv_gptj_rope(o_decode, pos_d, cos_sin, NOPE, ROPE) + o_grp_d = o_inv_d.reshape(1, OG, HPG * HD).permute(1, 0, 2) + z_d = torch.bmm(o_grp_d, woa_3d.transpose(1, 2)).permute(1, 0, 2).reshape(1, OG * OL) + attn_decode = r_wob.run(z_d) + + # ── REFERENCE: prefill all N+1 tokens, take the last ──── + all_ids = torch.cat([token_ids, decode_id]) + all_pos = torch.arange(N + 1, dtype=torch.int64, device=DEV) + hidden_ref = emb[all_ids] + normed_ref = rms(hidden_ref, anorm, EPS) + qa_ref = r_qa.run(normed_ref); kv_ref = r_kv.run(normed_ref) + qa_n_ref = rms(qa_ref, qn, EPS); kv_n_ref = rms(kv_ref, kvn, EPS) + q_ref = r_qb.run(qa_n_ref).view(N + 1, NH, HD) + q_rope_ref = apply_gptj_rope(q_ref, all_pos, cos_sin, NOPE, ROPE) + kv_rope_ref = apply_gptj_rope(kv_n_ref.unsqueeze(1), all_pos, cos_sin, NOPE, ROPE).squeeze(1) + o_ref = causal_prefill_attention(q_rope_ref, kv_rope_ref, SCALE) + o_inv_ref = apply_inv_gptj_rope(o_ref[-1:], pos_d, cos_sin, NOPE, ROPE) + o_grp_ref = o_inv_ref.reshape(1, OG, HPG * HD).permute(1, 0, 2) + z_ref = torch.bmm(o_grp_ref, woa_3d.transpose(1, 2)).permute(1, 0, 2).reshape(1, OG * OL) + attn_ref = r_wob.run(z_ref) + + # ── COMPARE ───────────────────────────────────────────── + # Decode attention output vs reference + c_attn = F.cosine_similarity(o_decode.flatten().unsqueeze(0).float(), o_ref[-1:].flatten().unsqueeze(0).float()).item() + # Full output vs reference + c_full = F.cosine_similarity(attn_decode.flatten().unsqueeze(0).float(), attn_ref.flatten().unsqueeze(0).float()).item() + + del r_qa, r_qb, r_kv, r_wob + torch.cuda.empty_cache() + _cache.clear() + + return c_attn, c_full + + +def main(): + print("=" * 70) + print(" DeepSeek-V4 Decode vs Prefill Consistency Test") + print(" Verifies KV cache produces same output as full prefill") + print("=" * 70) + + test_layers = [ + (0, "C128A"), + (1, "C4A"), + (2, "C4A"), + (30, "C4A"), + (60, "SWA"), + ] + + for layer_id, lt in test_layers: + c_attn, c_full = test_layer_decode_vs_prefill(layer_id) + status = "✅" if c_full >= 0.98 else "❌" + print(f" Layer {layer_id} ({lt}): attn={c_attn:.4f} full={c_full:.4f} {status}") + + print(f"\n{'='*70}") + print(f" If all layers pass (≥0.98), the KV cache pipeline is correct.") + print(f" The vLLM container should produce valid output.") + print(f"{'='*70}") + + +if __name__ == "__main__": + main()