File size: 8,701 Bytes
34f2e1c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
201
202
203
204
205
206
207
208
209
210
211
212
"""
CPU-only held-out evaluation for BitLoopLM.

Runs safely alongside a live GPU training job — only uses CPU + RAM.
Loads the weights currently saved to $CKPT (default: pytorch_model.bin
written by save_and_push at SAVE_INTERVAL milestones), streams a
held-out slice of the training dataset, and reports pure CE + perplexity.

Config via env vars (all optional):
  CKPT              pytorch_model.bin / resume.pt path (default: latest hub artifact)
  MODEL_SIZE        "small" | "tiny"   (must match the checkpoint)
  NUM_LOOPS         number of recurrent loops (must match)
  EVAL_DATASET      HF dataset name
  EVAL_DATASET_CONFIG  HF dataset config
  EVAL_SKIP         samples to discard before taking the eval window (ensures held-out)
  EVAL_BATCHES      number of evaluation batches
  EVAL_SEQ_LEN      tokens per sample (smaller = faster on CPU)
  EVAL_BATCH_SIZE   batch dim
  TOKENIZER         HF tokenizer id
  FAST              "1" (default) = pre-freeze BitLinear weights + bf16 cast.
                    "0" disables for an apples-to-apples baseline comparison.

Usage:
  uv run python eval_cpu.py
  CKPT=./bitlooplm-checkpoints/resume.pt EVAL_BATCHES=16 uv run python eval_cpu.py
  FAST=0 uv run python eval_cpu.py    # baseline (slower, fp32, no freeze)
"""
import math
import os
import sys
import time

import torch
import torch.nn.functional as F

# Import model from the training script (same directory)
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from train_bitlooplm_standalone import (
    BitLoopLM, BitLoopLMConfig, MODEL_CONFIGS,
    BitLinear, weight_quant, activation_quant,
)

CKPT = os.environ.get("CKPT", "./bitlooplm-checkpoints/pytorch_model.bin")
MODEL_SIZE = os.environ.get("MODEL_SIZE", "small")
NUM_LOOPS = int(os.environ.get("NUM_LOOPS", "4"))
TOKENIZER = os.environ.get("TOKENIZER", "HuggingFaceTB/SmolLM2-135M")
EVAL_DATASET = os.environ.get("EVAL_DATASET", "HuggingFaceTB/smollm-corpus")
EVAL_DATASET_CONFIG = os.environ.get("EVAL_DATASET_CONFIG", "cosmopedia-v2")
EVAL_SKIP = int(os.environ.get("EVAL_SKIP", "50000"))
EVAL_BATCHES = int(os.environ.get("EVAL_BATCHES", "8"))
EVAL_SEQ_LEN = int(os.environ.get("EVAL_SEQ_LEN", "256"))
EVAL_BATCH_SIZE = int(os.environ.get("EVAL_BATCH_SIZE", "4"))
FAST = os.environ.get("FAST", "1") == "1"


def freeze_bitlinears(model):
    """Pre-quantize BitLinear weights once, swap forward to skip per-call quant + STE.

    The training script's BitLinear recomputes weight_quant() on every forward
    (correct for STE) and pads it with detach-add tricks (training-only). For
    eval, weights are static, so we pay both costs for nothing. Walk the model,
    cache the quantized weight as a buffer (so .to(dtype) casts it along with
    everything else), and replace forward with a lean version.

    Returns the number of layers patched.
    """
    import types
    n = 0
    for module in model.modules():
        if isinstance(module, BitLinear):
            with torch.no_grad():
                qw = weight_quant(module.weight).detach().clone()
            module.register_buffer("_quant_weight_cached", qw, persistent=False)

            def fast_forward(self, x):
                qx = activation_quant(x)
                out = F.linear(qx, self._quant_weight_cached)
                if self.bias is not None:
                    out = out + self.bias
                return out

            module.forward = types.MethodType(fast_forward, module)
            n += 1
    return n


def main():
    torch.set_num_threads(max(1, (os.cpu_count() or 4) - 1))
    device = torch.device("cpu")

    print(f"[eval] device=cpu threads={torch.get_num_threads()}")
    print(f"[eval] loading tokenizer: {TOKENIZER}")
    from transformers import AutoTokenizer
    tokenizer = AutoTokenizer.from_pretrained(TOKENIZER)

    print(f"[eval] building model: size={MODEL_SIZE} loops={NUM_LOOPS}")
    cfg_dict = dict(MODEL_CONFIGS[MODEL_SIZE])
    cfg_dict["num_loops"] = NUM_LOOPS
    config = BitLoopLMConfig(**cfg_dict)
    model = BitLoopLM(config)
    model.eval()

    print(f"[eval] loading checkpoint: {CKPT}")
    state = torch.load(CKPT, map_location="cpu", weights_only=False)
    if isinstance(state, dict) and "model" in state:
        # resume.pt wraps weights under "model"
        state = state["model"]
    missing, unexpected = model.load_state_dict(state, strict=False)
    if missing:
        print(f"[eval] WARNING missing keys: {len(missing)} (e.g., {missing[:3]})")
    if unexpected:
        print(f"[eval] WARNING unexpected keys: {len(unexpected)} (e.g., {unexpected[:3]})")

    n_params = sum(p.numel() for p in model.parameters())
    print(f"[eval] model loaded, {n_params/1e6:.1f}M params")

    if FAST:
        # Tier 1+2: bf16 GEMM + pre-frozen BitLinear weights.
        torch.set_float32_matmul_precision("medium")
        n_frozen = freeze_bitlinears(model)
        # Cast AFTER freezing so weight_quant runs on fp32 weights, then the
        # cached quantized buffer is cast to bf16 along with everything else.
        model = model.to(torch.bfloat16)
        print(f"[eval] FAST mode: pre-froze {n_frozen} BitLinears, cast model to bf16")

    print(f"[eval] streaming {EVAL_DATASET}/{EVAL_DATASET_CONFIG}, skip={EVAL_SKIP}")
    from datasets import load_dataset
    ds = load_dataset(
        EVAL_DATASET, EVAL_DATASET_CONFIG,
        split="train", streaming=True,
    )
    ds = ds.skip(EVAL_SKIP)

    # Collect enough tokens for EVAL_BATCHES * EVAL_BATCH_SIZE * EVAL_SEQ_LEN
    needed = EVAL_BATCHES * EVAL_BATCH_SIZE * EVAL_SEQ_LEN
    print(f"[eval] need {needed} tokens for {EVAL_BATCHES} batches of {EVAL_BATCH_SIZE}x{EVAL_SEQ_LEN}")

    buffer = []
    samples_consumed = 0
    t0 = time.time()
    for sample in ds:
        text = sample.get("text") or sample.get("content") or ""
        if not text:
            continue
        ids = tokenizer.encode(text, add_special_tokens=False)
        buffer.extend(ids)
        samples_consumed += 1
        if len(buffer) >= needed:
            break
    print(f"[eval] collected {len(buffer)} tokens from {samples_consumed} samples in {time.time()-t0:.1f}s")

    total_ce = 0.0
    total_tokens = 0
    total_forward_time = 0.0
    total_loop_ce = torch.zeros(config.num_loops)
    total_exit = torch.zeros(config.num_loops)

    print(f"[eval] running {EVAL_BATCHES} batches")
    with torch.no_grad():
        for b in range(EVAL_BATCHES):
            off = b * EVAL_BATCH_SIZE * EVAL_SEQ_LEN
            chunk = buffer[off: off + EVAL_BATCH_SIZE * EVAL_SEQ_LEN]
            batch = torch.tensor(chunk, dtype=torch.long, device=device).view(
                EVAL_BATCH_SIZE, EVAL_SEQ_LEN
            )

            t0 = time.time()
            # Inference path: returns (weighted-combined logits, exit_pdf)
            logits, exit_pdf = model(batch)
            # Cast logits to fp32 for CE — bf16 softmax can underflow on the rare-token tail.
            shift_logits = logits[:, :-1, :].contiguous().float()
            shift_labels = batch[:, 1:].contiguous()
            ce_tokens = F.cross_entropy(
                shift_logits.view(-1, config.vocab_size),
                shift_labels.view(-1),
                reduction="sum",
            )
            total_ce += ce_tokens.item()
            total_tokens += shift_labels.numel()
            total_exit += exit_pdf.mean(dim=(0, 1)).detach()
            dt = time.time() - t0
            total_forward_time += dt

            # Per-loop CE (inference-mode, unweighted, for diagnostics)
            with torch.no_grad():
                # Re-run in labels mode to get per-loop stats without extra grad cost
                pass  # skip — would double the forward cost. Use training logs for per-loop.

            avg = total_ce / total_tokens
            print(
                f"  batch {b+1}/{EVAL_BATCHES}: ce={ce_tokens.item()/shift_labels.numel():.4f}  "
                f"running_avg={avg:.4f}  ppl={math.exp(avg):.2f}  "
                f"forward={dt:.1f}s"
            )

    avg_ce = total_ce / total_tokens
    ppl = math.exp(avg_ce)
    exit_mean = (total_exit / EVAL_BATCHES).tolist()

    print("\n=== Eval summary ===")
    print(f"  checkpoint      : {CKPT}")
    print(f"  held-out dataset: {EVAL_DATASET}/{EVAL_DATASET_CONFIG}  skip={EVAL_SKIP}")
    print(f"  tokens evaluated: {total_tokens}")
    print(f"  avg CE (nats)   : {avg_ce:.4f}")
    print(f"  perplexity      : {ppl:.2f}")
    print(f"  exit pdf        : {[f'L{i}:{v:.2f}' for i, v in enumerate(exit_mean)]}")
    print(f"  forward time    : {total_forward_time:.1f}s total ({total_forward_time/EVAL_BATCHES:.1f}s/batch)")


if __name__ == "__main__":
    main()