headroom/scripts/export_kompress_v2_onnx.py
Tejas Chopra 74392b238e
feat: switch Kompress default to kompress-v2-base with weight-only int8 ONNX (#799)
## Summary

Replaces `chopratejas/kompress-base` with
**`chopratejas/kompress-v2-base`** as the default Kompress
text-compression model (the fallback for content not handled by
structured compressors), using a new **weight-only int8 ONNX** artifact
that is fp32-equivalent at 2.2x less memory.

## Why

v2 is the same dual-head ModernBERT (token classifier + span CNN),
LoRA-trained on the v2 dataset. The HF repo originally shipped PyTorch
weights only — pointing Headroom at it naively would have forced the
heavy `[ml]` (torch) path on every `[proxy]` install. We exported ONNX
artifacts reproducing the v1 loader contract (single `final_scores`
output) and published them to the HF repo.

## Eval (labeled dataset_v2 test split, n=500, threshold 0.5)

| artifact | size | f1 | must_keep_recall | keep_rate | fp32 agreement |
|---|---|---|---|---|---|
| fp32 (reference) | 601 MB | 0.9128 | 0.9770 | 0.8100 | 100% |
| **int8-wo (default)** | **261 MB** | **0.9130** | **0.9765** |
**0.8097** | **99.6%** |
| fp16 | 301 MB | 0.9129 | 0.9771 | 0.8099 | 99.9% |
| int4-wo | 230 MB | 0.9102 | 0.9825 | 0.8354 ⚠️ | 94.5% |
| int8-dynamic | 271 MB | 0.9041 | 0.9868 | 0.8857 ⚠️ | 90.5% |

Weight-only int8 (MatMulNBits) keeps activations fp32, avoiding the
upward score bias that makes dynamic int8 keep ~7% more tokens (≈40%
less compression savings). Quantized candidates were generated and
eval-gated by a Modal job in the kompress repo
(`modal_jobs/export_onnx_v2.py`) against the labeled test split.

## Changes

- Default model id → `chopratejas/kompress-v2-base`
- ONNX artifact resolution tries candidates in order (**int8-wo → fp32 →
v1 int8**), falling through on download miss **or session-load failure**
— onnxruntime builds without the MatMulNBits 8-bit kernel fall back to
fp32 instead of losing Kompress
- `HEADROOM_KOMPRESS_ONNX_FILENAME` env pins an exact artifact
- `scripts/export_kompress_v2_onnx.py`: reproducible fp32 export (loads
the merged v2 checkpoint, traces the `final_scores` contract, verifies
vs PyTorch)
- `.gitignore`: local `onnx/` artifacts dir; allowlist the export script

## Testing

- End-to-end: fresh `KompressCompressor().preload()` resolves int8-wo
from HF, loads on onnxruntime 1.23 (CPU, no torch), compresses with
error/traceback content preserved
- fp32 export verified lossless vs PyTorch (max |Δscore| = 0.0, 100%
keep agreement)
- ruff check + format clean, mypy clean, 63 targeted tests pass
2026-06-09 23:28:40 -07:00

267 lines
9.9 KiB
Python

#!/usr/bin/env python
"""Export a Kompress PyTorch checkpoint to ONNX INT8 for Headroom's light path.
Why this exists
---------------
Headroom's ``[proxy]`` extra ships ``onnxruntime`` but **not** torch — the
proxy runs Kompress text compression on ONNX Runtime alone. The loader
(``headroom/transforms/kompress_compressor.py``) downloads
``onnx/kompress-int8.onnx`` from the model repo and runs it through
``_OnnxModel``, which expects a single graph output named ``final_scores``
(per-token importance in ``[0, 1]``, kept when ``> 0.5``).
``chopratejas/kompress-v2-base`` ships only PyTorch weights
(``model.safetensors`` / ``merged.pt``) — no ONNX. So pointing Headroom at v2
without an ONNX export would silently force the heavier ``[ml]`` (torch) path
on every proxy install. This script reproduces v1's exact ONNX contract from
the v2 PyTorch checkpoint, so a default swap stays zero-cost for light installs.
The model is a *custom* dual-head ModernBERT (token classifier + span CNN), not
a standard HF architecture, so ``optimum-cli export onnx`` does not apply — we
trace the real module from ``kompress_compressor._get_model_class()``.
Requires
--------
pip install headroom-ai[ml] onnxruntime # torch + transformers + onnxruntime
Usage
-----
# Convert + verify locally (writes onnx/kompress-int8.onnx):
python scripts/export_kompress_v2_onnx.py --model-id chopratejas/kompress-v2-base
# Convert, verify, and upload back to the HF repo (needs `huggingface-cli login`):
python scripts/export_kompress_v2_onnx.py --model-id chopratejas/kompress-v2-base --upload
"""
from __future__ import annotations
import argparse
import logging
import sys
from pathlib import Path
logging.basicConfig(level=logging.INFO, format="%(levelname)s %(message)s")
logger = logging.getLogger("export_kompress_v2_onnx")
# ModernBERT encoder + tokenizer base (must match training and the loader).
BASE_MODEL = "answerdotai/ModernBERT-base"
DEFAULT_MODEL_ID = "chopratejas/kompress-v2-base"
def _build_core(model_id: str):
"""Instantiate HeadroomCompressorModel and load the merged v2 weights.
The v2 repo's ``model.safetensors`` is the *unmerged* PEFT structure
(``encoder.base_model.model...`` with separate ``base_layer`` + LoRA
adapters), which does not map onto ``HeadroomCompressorModel``. The
canonical artifact is ``merged.pt`` — a structured checkpoint with already
LoRA-merged sub-state-dicts:
{"encoder_state_dict", "token_head_state_dict",
"span_conv_state_dict", "config", "checkpoint_kind"}
Each loads cleanly (0 missing / 0 unexpected) into the encoder + heads.
"""
import torch
from huggingface_hub import hf_hub_download
from headroom.transforms.kompress_compressor import _get_model_class
ckpt_path = hf_hub_download(model_id, "merged.pt")
ckpt = torch.load(ckpt_path, map_location="cpu")
for key in ("encoder_state_dict", "token_head_state_dict", "span_conv_state_dict"):
if key not in ckpt:
raise RuntimeError(
f"merged.pt missing '{key}'. Found: {sorted(ckpt)}. "
"This script targets the v2 'merged' checkpoint format."
)
core = _get_model_class()(model_name=BASE_MODEL)
def _strict_load(module, sd, label: str) -> None:
missing, unexpected = module.load_state_dict(sd, strict=False)
if missing or unexpected:
raise RuntimeError(
f"{label}: state_dict mismatch (missing={list(missing)[:5]}, "
f"unexpected={list(unexpected)[:5]}). Architecture drifted from the checkpoint."
)
logger.info(" %s loaded (%d tensors, exact match)", label, len(sd))
logger.info("Loading merged.pt (checkpoint_kind=%s)", ckpt.get("checkpoint_kind"))
_strict_load(core.encoder, ckpt["encoder_state_dict"], "encoder")
_strict_load(core.token_head, ckpt["token_head_state_dict"], "token_head")
_strict_load(core.span_conv, ckpt["span_conv_state_dict"], "span_conv")
core.eval()
return core
def _export_wrapper(core):
"""Wrap the dual head so forward() returns `final_scores` (== get_scores)."""
import torch
import torch.nn as nn
class ExportWrapper(nn.Module):
def __init__(self, inner):
super().__init__()
self.inner = inner
def forward(self, input_ids, attention_mask): # noqa: ANN001
hidden = self.inner.encoder(input_ids, attention_mask=attention_mask).last_hidden_state
token_probs = torch.softmax(self.inner.token_head(hidden), dim=-1)[:, :, 1]
span_scores = self.inner.span_conv(hidden.transpose(1, 2)).squeeze(1)
return token_probs * (0.5 + 0.5 * span_scores)
return ExportWrapper(core).eval()
def export(model_id: str, out_path: Path, opset: int, precision: str) -> None:
import numpy as np
import torch
core = _build_core(model_id)
wrapper = _export_wrapper(core)
out_path.parent.mkdir(parents=True, exist_ok=True)
# fp32 path: trace straight to the final artifact (lossless — verified 100%
# keep-decision agreement with PyTorch). int8 path: trace to a temp fp32
# graph, then dynamically quantize into the final artifact.
trace_target = out_path if precision == "fp32" else out_path.with_name("kompress-fp32-tmp.onnx")
dummy_ids = torch.randint(0, 1000, (1, 64), dtype=torch.long)
dummy_mask = torch.ones((1, 64), dtype=torch.long)
logger.info("Tracing → ONNX (opset %d, precision=%s) ...", opset, precision)
with torch.no_grad():
torch.onnx.export(
wrapper,
(dummy_ids, dummy_mask),
str(trace_target),
input_names=["input_ids", "attention_mask"],
output_names=["final_scores"],
dynamic_axes={
"input_ids": {0: "batch", 1: "seq"},
"attention_mask": {0: "batch", 1: "seq"},
"final_scores": {0: "batch", 1: "seq"},
},
opset_version=opset,
do_constant_folding=True,
dynamo=False,
)
if precision == "int8":
from onnxruntime.quantization import QuantType, quantize_dynamic
logger.info("INT8 dynamic quantization (MatMul only) → %s", out_path)
# Restrict to MatMul: the encoder's linear layers carry ~all the weight
# mass and ORT's CPU provider implements MatMulInteger. Quantizing the
# tiny span_conv Conv1d layers would emit ConvInteger, which ORT CPU
# cannot run. per_channel recovers transformer accuracy at the 0.5 boundary.
quantize_dynamic(
str(trace_target),
str(out_path),
weight_type=QuantType.QInt8,
op_types_to_quantize=["MatMul"],
per_channel=True,
)
trace_target.unlink(missing_ok=True)
_verify(model_id, core, out_path, np, torch)
def _verify(model_id: str, core, out_path: Path, np, torch) -> None:
"""Compare ONNX scores against PyTorch get_scores on a real tokenized sample."""
import onnxruntime as ort
from transformers import AutoTokenizer
tok = AutoTokenizer.from_pretrained(BASE_MODEL)
sample = (
"The proxy compresses tool outputs before they reach the model. "
"Errors and stack traces should survive; boilerplate should not. "
) * 6
words = sample.split()
enc = tok(
words,
is_split_into_words=True,
truncation=True,
max_length=512,
padding=True,
return_tensors="pt",
)
with torch.no_grad():
torch_scores = core.get_scores(enc["input_ids"], enc["attention_mask"])[0].cpu().numpy()
sess = ort.InferenceSession(str(out_path), providers=["CPUExecutionProvider"])
onnx_scores = sess.run(
["final_scores"],
{
"input_ids": enc["input_ids"].numpy().astype(np.int64),
"attention_mask": enc["attention_mask"].numpy().astype(np.int64),
},
)[0][0]
max_abs = float(np.max(np.abs(torch_scores - onnx_scores)))
keep_torch = torch_scores > 0.5
keep_onnx = onnx_scores > 0.5
agree = float((keep_torch == keep_onnx).mean())
logger.info(
"Verify: max|Δscore|=%.4f keep-decision agreement=%.1f%% (fp32 ~100%%, int8 ~98-100%%)",
max_abs,
agree * 100,
)
if agree < 0.98:
logger.warning(
"Keep-decision agreement below 98%% — for fp32 this means a tracing "
"problem; for int8 consider per_channel/fp32. Inspect before publishing."
)
def upload(model_id: str, out_path: Path) -> None:
from huggingface_hub import upload_file
# Publish under onnx/<artifact filename> so int8 and fp32 can coexist.
repo_path = f"onnx/{out_path.name}"
logger.info("Uploading %s%s:%s", out_path, model_id, repo_path)
upload_file(
path_or_fileobj=str(out_path),
path_in_repo=repo_path,
repo_id=model_id,
commit_message="Add ONNX export for Headroom lightweight (no-torch) path",
)
logger.info("Uploaded. Headroom's ONNX loader will now find it on next cold start.")
def main() -> int:
ap = argparse.ArgumentParser(description=__doc__)
ap.add_argument("--model-id", default=DEFAULT_MODEL_ID)
ap.add_argument(
"--precision",
choices=["fp32", "int8"],
default="fp32",
help="fp32 = lossless, larger artifact. int8 = ~2x smaller, tiny accuracy cost.",
)
ap.add_argument(
"--out",
type=Path,
default=None,
help="Local output path. Defaults to onnx/kompress-<precision>.onnx.",
)
ap.add_argument("--opset", type=int, default=17)
ap.add_argument(
"--upload",
action="store_true",
help="Upload to the HF repo under onnx/<filename> (needs HF write auth).",
)
args = ap.parse_args()
out_path = args.out or Path(f"onnx/kompress-{args.precision}.onnx")
export(args.model_id, out_path, args.opset, args.precision)
if args.upload:
upload(args.model_id, out_path)
return 0
if __name__ == "__main__":
sys.exit(main())