File size: 8,323 Bytes
b19c63d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
213
214
215
216
217
#!/usr/bin/env python3
"""
Re-export zerank-1-small with dynamic batch support.

Key change from v1: ZeRankScorerV2 builds the 4D causal+padding attention mask
explicitly using input_ids.shape[0] (dynamic). This makes the batch dimension
symbolic in the ONNX graph — batch > 1 works correctly.

Also bakes the Qwen3 chat template into the expected input format:
  "<|im_start|>user\\nQuery: {q}\\nDocument: {d}\\nRelevant:<|im_end|>\\n<|im_start|>assistant\\n"

Tokenize the formatted string as a SINGLE sequence (not a pair) in fastembed.

Output:
  /private/tmp/zerank_export/zerank_onnx_v2/model.onnx + model.onnx_data  (FP16)
  (INT8/INT4 re-quantization: run stream_int8.py and export_int4.py after this)
"""

import gc
from pathlib import Path
import numpy as np
import torch
import torch.nn as nn

MODEL_ID  = "zeroentropy/zerank-1-small"
YES_TOKEN_ID = 9454

OUT_DIR   = Path("/private/tmp/zerank_export/zerank_onnx_v2")
OUT_MODEL = OUT_DIR / "model.onnx"
OUT_DIR.mkdir(parents=True, exist_ok=True)


class ZeRankScorerV2(nn.Module):
    """
    Wraps Qwen3ForCausalLM + last-token Yes-logit extraction.

    Difference from V1: builds 4D causal+padding mask explicitly so the batch
    dimension is dynamic in the ONNX graph (V1 had it hardcoded to 1).

    Input:
      input_ids      [batch, seq] — pre-formatted with chat template
      attention_mask [batch, seq] — 1 for real tokens, 0 for padding

    Output:
      logits [batch, 1] — raw Yes-token logit, higher = more relevant
    """
    def __init__(self, base_model, yes_token_id: int):
        super().__init__()
        self.base = base_model
        self.yes_token_id = yes_token_id
        self._dtype = next(base_model.parameters()).dtype

    def forward(self, input_ids: torch.Tensor, attention_mask: torch.Tensor):
        batch_size = input_ids.shape[0]
        seq_len    = input_ids.shape[1]
        device     = input_ids.device
        min_val    = torch.finfo(self._dtype).min

        # Causal mask: upper-triangular = min_val, lower-triangular = 0
        # Shape [1, 1, seq, seq] → expand to [batch, 1, seq, seq]
        upper = torch.ones(seq_len, seq_len, dtype=torch.bool, device=device).triu(diagonal=1)
        causal = torch.zeros(1, 1, seq_len, seq_len, dtype=self._dtype, device=device)
        causal = causal.masked_fill(upper.view(1, 1, seq_len, seq_len), min_val)
        causal = causal.expand(batch_size, 1, seq_len, seq_len)

        # Padding mask: positions with attention_mask=0 get min_val
        pad = (1.0 - attention_mask.to(self._dtype)) * min_val  # [batch, seq]
        pad = pad.unsqueeze(1).unsqueeze(2)                     # [batch, 1, 1, seq]
        pad = pad.expand(batch_size, 1, seq_len, seq_len)

        full_mask = causal + pad

        # Transformer body → [batch, seq, hidden]
        hidden = self.base.model(
            input_ids=input_ids,
            attention_mask=full_mask,
        )[0]

        # Gather at last real-token position: sum(mask) - 1
        last_pos = attention_mask.sum(dim=-1) - 1  # [batch]
        idx = last_pos.view(-1, 1, 1).expand(-1, 1, hidden.shape[-1])
        last_hidden = torch.gather(hidden, 1, idx).squeeze(1)  # [batch, hidden]

        yes_logit = self.base.lm_head(last_hidden)[:, self.yes_token_id]  # [batch]
        return yes_logit.unsqueeze(-1)  # [batch, 1]


def run_export():
    from transformers import Qwen3ForCausalLM, AutoTokenizer
    import torch.onnx as torch_onnx

    print(f"Loading {MODEL_ID}...")
    model = Qwen3ForCausalLM.from_pretrained(
        MODEL_ID,
        torch_dtype=torch.float16,
        low_cpu_mem_usage=True,
        attn_implementation="eager",
    ).eval()

    scorer = ZeRankScorerV2(model, YES_TOKEN_ID).eval()

    tok = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)

    # Dummy batch=2 — forces dynamic batch to trace correctly
    template = "<|im_start|>user\nQuery: {q}\nDocument: {d}\nRelevant:<|im_end|>\n<|im_start|>assistant\n"
    pairs = [
        ("what is a panda?", "A panda is a large black-and-white bear."),
        ("what is a cat?",   "A cat is a small domesticated carnivorous mammal."),
    ]
    formatted = [template.format(q=q, d=d) for q, d in pairs]
    enc = tok(formatted, padding=True, truncation=True, max_length=64, return_tensors="pt")
    dummy_ids  = enc["input_ids"]
    dummy_mask = enc["attention_mask"]
    print(f"  Dummy batch shape: {dummy_ids.shape}")

    # Verify correct batch behaviour before exporting
    with torch.no_grad():
        out_batch = scorer(dummy_ids, dummy_mask)
        out_single = scorer(dummy_ids[:1], dummy_mask[:1])
    assert abs(float(out_batch[0, 0]) - float(out_single[0, 0])) < 0.01, \
        f"Batch/single mismatch: {float(out_batch[0,0]):.3f} vs {float(out_single[0,0]):.3f}"
    print(f"  Batch consistency check PASS: {float(out_batch[0,0]):.3f} vs {float(out_single[0,0]):.3f}")

    print(f"Exporting to {OUT_MODEL} ...")
    with torch.no_grad():
        torch_onnx.export(
            scorer,
            (dummy_ids, dummy_mask),
            str(OUT_MODEL),
            input_names=["input_ids", "attention_mask"],
            output_names=["logits"],
            dynamic_axes={
                "input_ids":      {0: "batch_size", 1: "sequence_length"},
                "attention_mask": {0: "batch_size", 1: "sequence_length"},
                "logits":         {0: "batch_size"},
            },
            opset_version=18,
            do_constant_folding=False,
        )

    import onnx
    from onnx.external_data_helper import convert_model_to_external_data
    print("  Converting to external data format...")
    m = onnx.load(str(OUT_MODEL))
    convert_model_to_external_data(
        m, all_tensors_to_one_file=True,
        location="model.onnx_data", size_threshold=1024,
    )
    onnx.save(m, str(OUT_MODEL))
    print("Export complete:")
    for f in sorted(OUT_DIR.iterdir()):
        print(f"  {f.name:40s} {f.stat().st_size / 1e6:.0f} MB")

    del m, scorer, model, tok, enc, dummy_ids, dummy_mask
    gc.collect()


def verify_batch():
    import onnxruntime as ort

    print(f"\nVerifying batch > 1...")
    sess = ort.InferenceSession(str(OUT_MODEL), providers=["CPUExecutionProvider"])

    from transformers import AutoTokenizer
    tok = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
    template = "<|im_start|>user\nQuery: {q}\nDocument: {d}\nRelevant:<|im_end|>\n<|im_start|>assistant\n"

    q = "what is a panda?"
    docs = [
        "The giant panda is a bear species endemic to China.",
        "The sky is blue.",
        "panda is an animal",
    ]

    # Single inference
    single_scores = []
    for d in docs:
        fmt = template.format(q=q, d=d)
        enc = tok(fmt, return_tensors="np", truncation=True, max_length=256)
        logit = sess.run(["logits"], {
            "input_ids":      enc["input_ids"].astype(np.int64),
            "attention_mask": enc["attention_mask"].astype(np.int64),
        })[0]
        single_scores.append(float(logit[0, 0]))

    # Batch inference
    formatted = [template.format(q=q, d=d) for d in docs]
    enc = tok(formatted, return_tensors="np", truncation=True, max_length=256, padding=True)
    logits = sess.run(["logits"], {
        "input_ids":      enc["input_ids"].astype(np.int64),
        "attention_mask": enc["attention_mask"].astype(np.int64),
    })[0]
    batch_scores = [float(logits[i, 0]) for i in range(len(docs))]

    print("  Single vs batch scores:")
    for d, s, b in zip(docs, single_scores, batch_scores):
        diff = abs(s - b)
        print(f"  [{s:.3f} vs {b:.3f}] diff={diff:.4f} | {d[:50]}")
        assert diff < 0.1, f"Mismatch too large: {diff}"
    assert batch_scores[0] > batch_scores[1], "Panda should rank higher than sky"
    print("  OK — batch scores match single, correct ranking")


if __name__ == "__main__":
    if OUT_MODEL.exists():
        print(f"Model already exists at {OUT_MODEL}, skipping export.")
        print("Delete it to re-export.")
    else:
        run_export()
        gc.collect()

    verify_batch()

    print("\nNext steps:")
    print(f"  1. Run stream_int8_v2.py to quantize INT8 from {OUT_MODEL}")
    print(f"  2. Upload to HF: huggingface-cli upload cstr/zerank-1-small-ONNX {OUT_DIR}/ . --repo-type model")