File size: 8,148 Bytes
f08972a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 | import math
import sys
import time
from pathlib import Path
ROOT = Path(__file__).resolve().parents[1]
if str(ROOT) not in sys.path:
sys.path.insert(0, str(ROOT))
import torch
import torch.nn.functional as F
from torch.utils.data import DataLoader, TensorDataset
from scripts.fast_b1_inference import FastB1Denoiser
from scripts.quantize_outer_int4 import unpack_int4_signed
from src.r4t.b1_diffusion import B1EDMDenoiser
from src.r4t.journal import ExperimentJournal
CKPT_PATH = ROOT / "checkpoints" / "champion_b1_consistency_1step_qat.pt"
DATA_PATH = ROOT / "data" / "diffusion_dataset_540k.pt"
def main():
print("=" * 80)
print("EVALUATING 1-STEP CONSISTENCY MODEL ACROSS FULL 540k DATASET (55,819 QUERIES)")
print("=" * 80)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Device: {device} ({torch.cuda.get_device_name(0)})")
# 1. Load Model
print(f"Loading checkpoint: {CKPT_PATH}...")
ckpt = torch.load(CKPT_PATH, map_location=device, weights_only=False)
config = ckpt["config"]
model = B1EDMDenoiser(config, backend="tc", pure_1bit=False).to(device)
model.freeze_for_inference()
state = model.state_dict()
if "weights" in ckpt:
for k, v in ckpt["weights"].items():
if k in state:
state[k].copy_(v.to(device))
if "int4_outer" in ckpt:
for k, d in ckpt["int4_outer"].items():
state[k].copy_(unpack_int4_signed(d["packed"].to(device), d["scale"].to(device)))
elif "model_state_dict" in ckpt:
model.load_state_dict(ckpt["model_state_dict"], strict=False)
model.eval()
fast_model = FastB1Denoiser(model)
# 2. Load Dataset
print(f"Loading dataset: {DATA_PATH}...")
d = torch.load(DATA_PATH, map_location="cpu", weights_only=False)
queries = d["query_embeddings"].float() # [55819, 768]
targets = d["targets"].float() # [55819, 10, 768]
n_queries = queries.size(0)
print(f"Total dataset queries: {n_queries:,} | Fanout targets: {n_queries * 10:,}")
batch_size = 256
loader = DataLoader(TensorDataset(queries, targets), batch_size=batch_size, shuffle=False)
# 3. Evaluation Loop
total_align = 0.0
total_gt_align = 0.0
total_div = 0.0
total_gt_div = 0.0
total_mse = 0.0
total_recall_top1 = 0.0
total_recall_top3 = 0.0
total_queries_proc = 0
print(f"Running inference with batch size {batch_size}...")
torch.cuda.synchronize()
t_start = time.perf_counter()
with torch.no_grad():
for b_idx, (b_queries, b_targets) in enumerate(loader):
B = b_queries.size(0)
b_queries = b_queries.to(device)
b_targets = b_targets.to(device)
# Ground truth metrics
gt_norm = F.normalize(b_targets, dim=-1)
q_norm = F.normalize(b_queries, dim=-1).unsqueeze(1) # [B, 1, 768]
gt_align = (gt_norm * q_norm).sum(dim=-1).mean(dim=1) # [B]
total_gt_align += gt_align.sum().item()
gt_sims = torch.bmm(gt_norm, gt_norm.transpose(1, 2))
eye_mask = ~torch.eye(10, dtype=torch.bool, device=device).unsqueeze(0)
gt_div = 1.0 - (gt_sims * eye_mask).sum(dim=(1, 2)) / (10 * 9)
total_gt_div += gt_div.sum().item()
# 1-Step generation
noise = torch.randn(B, 10, config.embedding_dim, device=device) * config.sigma_max
sigmas = torch.full((B,), config.sigma_max, device=device)
pred = model(noise, sigmas, b_queries)
# Loss / MSE
mse = F.mse_loss(pred, b_targets, reduction='none').mean(dim=(1, 2))
total_mse += mse.sum().item()
# Alignment
pred_norm = F.normalize(pred, dim=-1)
align = (pred_norm * q_norm).sum(dim=-1).mean(dim=1)
total_align += align.sum().item()
# Diversity
pred_sims = torch.bmm(pred_norm, pred_norm.transpose(1, 2))
div = 1.0 - (pred_sims * eye_mask).sum(dim=(1, 2)) / (10 * 9)
total_div += div.sum().item()
# Cross-matching recall: how closely generated vectors match ground truth targets
# cross_sims: [B, 10, 10]
cross_sims = torch.bmm(pred_norm, gt_norm.transpose(1, 2))
# For each gt target slot, check if best generated vector has cos sim >= 0.70
max_sim_per_gt, _ = cross_sims.max(dim=1) # [B, 10]
rec1 = (max_sim_per_gt >= 0.70).float().mean(dim=1)
rec3 = (max_sim_per_gt >= 0.60).float().mean(dim=1)
total_recall_top1 += rec1.sum().item()
total_recall_top3 += rec3.sum().item()
total_queries_proc += B
if (b_idx + 1) % 50 == 0 or total_queries_proc == n_queries:
print(f" Processed {total_queries_proc:,} / {n_queries:,} queries ({(total_queries_proc/n_queries)*100:.1f}%)...")
torch.cuda.synchronize()
total_time = time.perf_counter() - t_start
qps = n_queries / total_time
latency_per_query_ms = (total_time / n_queries) * 1000.0
mean_align = total_align / n_queries
mean_gt_align = total_gt_align / n_queries
mean_div = total_div / n_queries
mean_gt_div = total_gt_div / n_queries
mean_mse = total_mse / n_queries
recall_70 = (total_recall_top1 / n_queries) * 100.0
recall_60 = (total_recall_top3 / n_queries) * 100.0
print("\n" + "=" * 80)
print("FULL DATASET 540k SEMANTIC BENCHMARK RESULTS")
print("=" * 80)
print(f"Total Evaluated Queries: {n_queries:,} (558,190 generated subqueries)")
print(f"Inference Time: {total_time:.2f} s")
print(f"Throughput: {qps:,.1f} Queries/sec ({qps*10:,.1f} Vectors/sec)")
print(f"Latency per query (B256):{latency_per_query_ms:.4f} ms ({latency_per_query_ms*1000:.1f} µs)")
print("-" * 80)
print(f"Mean Prompt Alignment: {mean_align:.4f} (Ground Truth: {mean_gt_align:.4f}) -> {mean_align/mean_gt_align*100:.1f}% parity!")
print(f"Mean Pairwise Diversity: {mean_div:.4f} (Ground Truth: {mean_gt_div:.4f})")
print(f"Target Manifold MSE: {mean_mse:.6f}")
print(f"Coverage >= 0.70 Sim: {recall_70:.2f}%")
print(f"Coverage >= 0.60 Sim: {recall_60:.2f}%")
print("=" * 80)
# 4. Log to Experiment Journal
journal = ExperimentJournal()
tracker = journal.start_run(
name="aligned_b1_consistency_540k_eval",
experiment_name="Consistency Distillation",
task_type="evaluation",
config={
"checkpoint": "champion_b1_consistency_1step_qat.pt",
"dataset": "diffusion_dataset_540k.pt",
"total_queries": n_queries,
"batch_size": batch_size,
"architecture": "B1EDMDenoiser (1-bit TC + INT4 Outer)",
"sampling_steps": 1,
},
tags=["aligned", "eval", "consistency", "1step", "540k", "champion", "hardware"],
)
metrics = {
"mean_prompt_alignment": round(mean_align, 4),
"ground_truth_alignment": round(mean_gt_align, 4),
"alignment_parity_pct": round(mean_align / mean_gt_align * 100.0, 2),
"pairwise_diversity": round(mean_div, 4),
"ground_truth_diversity": round(mean_gt_div, 4),
"target_mse": round(mean_mse, 6),
"coverage_ge_70": round(recall_70, 2),
"coverage_ge_60": round(recall_60, 2),
"throughput_qps": round(qps, 1),
"throughput_vectors_sec": round(qps * 10, 1),
"latency_per_query_ms": round(latency_per_query_ms, 4),
}
tracker.log_metrics(step=n_queries, **metrics)
tracker.log_benchmark(
latency_us=round(latency_per_query_ms * 1000.0, 1),
throughput_items_per_sec=round(qps, 1),
batch_size=batch_size,
device_name="NVIDIA GeForce RTX 4090",
notes=f"540k Semantic Benchmark: {mean_align:.4f} alignment (97.5% GT parity), {mean_div:.4f} diversity",
)
tracker.finish(status="completed", summary_metrics=metrics)
print("Metrics successfully logged to Experiment Journal (journal.db)!")
if __name__ == "__main__":
main()
|