Skip to content

Open-source AI browser agent for Chrome and Firefox (monorepo) 🧠

Tighten prefill evidence checks and make bundle replacement crash-safe

Claude4dfef545634c
diff --git a/scripts/models/bonsai/build_vision.py b/scripts/models/bonsai/build_vision.py
index 05f0b9ecb..4d40df926 100644
--- a/scripts/models/bonsai/build_vision.py
+++ b/scripts/models/bonsai/build_vision.py
@@ -339,7 +339,7 @@ optional check. {ocr_limit}
Apache 2.0; see LICENSE and NOTICE.txt. Original context capacity is 262,144 tokens;
WebBrain applies a smaller browser context budget. Model weights and processor
files are downloaded once; inference runs locally.
-""" + model_card_section([], language_graph["output_sha256"]), encoding="utf-8")
+""" + model_card_section([], language_graph["output_sha256"], language_graph["chunk_tokens"]), encoding="utf-8")
total = seal_manifest(output)
print(json.dumps({"output": str(output.resolve()), "vision_tensors": len(records),
"total_bytes": total, "vision_inference": validation}), flush=True)
diff --git a/scripts/models/bonsai/chunk_gdn_prefill.py b/scripts/models/bonsai/chunk_gdn_prefill.py
index 6af975b8e..6c2ec23a2 100644
--- a/scripts/models/bonsai/chunk_gdn_prefill.py
+++ b/scripts/models/bonsai/chunk_gdn_prefill.py
@@ -50,7 +50,7 @@ kernel, so prompts were processed one token at a time through a CPU loop with
a GPU round trip per token and layer (0.13 s per prompt token for 2B and 0.19 s
for 4B on an RTX 5090; a faster GPU barely helps).
This graph replaces each `Scan` with the chunked gated delta rule
-(128-token chunks; stable recursive block inversion inside a chunk; a parallel
+({chunk}-token chunks; stable recursive block inversion inside a chunk; a parallel
prefix scan across chunks), built only from ops the WebGPU EP implements, and
returns logits for the last position only. Weights, quantization and the
external data file are byte-identical to the previous revision; decode
@@ -63,6 +63,11 @@ MAX_RELATIVE_LOGIT_DIFF = 5e-3
MAX_ABSOLUTE_LOGIT_DIFF = 0.05
MAX_KL = 1e-3
MAX_WEBGPU_KL = 1e-2
+# Cache outputs (recurrent/conv states, KV) carry FP16 rounding through the
+# stack: observed whole-tensor error stays below 1% and the worst single
+# element below 3% of the tensor maximum. A corrupted cache is O(1).
+MAX_CACHE_REL_FRO = 2e-2
+MAX_CACHE_REL_MAX = 1e-1
def _require(condition, message):
@@ -88,6 +93,12 @@ def verify_prefill_evidence(reports, graph_sha256):
"CPU report must cover a chunk boundary and a multi-chunk prompt")
for row in rows + [continuation]:
_require(row.get("top1_equal") is True, f"top-1 differs ({row})")
+ _require(row.get("finite") is True, f"non-finite outputs ({row})")
+ _require("worst_cache_rel_fro" in row and "worst_cache_rel_err" in row,
+ "CPU report must record cache-output errors (rerun validate_chunked_gdn.py)")
+ _require(float(row["worst_cache_rel_fro"]) <= MAX_CACHE_REL_FRO
+ and float(row["worst_cache_rel_err"]) <= MAX_CACHE_REL_MAX,
+ f"cache outputs diverge ({row})")
# Without a recorded scale, allow a few FP16 ulps at typical |logit| of 16-32.
tolerance = (MAX_RELATIVE_LOGIT_DIFF * max(float(row["last_logits_scale"]), 1.0)
if row.get("last_logits_scale") else MAX_ABSOLUTE_LOGIT_DIFF)
@@ -97,8 +108,7 @@ def verify_prefill_evidence(reports, graph_sha256):
for row in rows:
_require(row.get("decode_step_top1_equal") is True, f"decode step differs ({row})")
elif kind == WEBGPU_EVIDENCE_KIND:
- if "rewritten_graph_sha256" in report:
- _require(report["rewritten_graph_sha256"] == graph_sha256, "WebGPU report is for a different graph")
+ _require(report.get("rewritten_graph_sha256") == graph_sha256, "WebGPU report is for a different graph")
rewritten = [row for row in report.get("prefill") or [] if row.get("variant") == "rewritten"]
original = [row for row in report.get("prefill") or [] if row.get("variant") == "original"]
comparisons = list(report.get("webgpu_logit_comparison") or [])
@@ -112,9 +122,9 @@ def verify_prefill_evidence(reports, graph_sha256):
return verified
-def model_card_section(reports, graph_sha256):
+def model_card_section(reports, graph_sha256, chunk_tokens=128):
verified = verify_prefill_evidence(reports, graph_sha256)
- text = MODEL_CARD_INTRO + "\n"
+ text = MODEL_CARD_INTRO.format(chunk=chunk_tokens) + "\n"
if CPU_EVIDENCE_KIND in verified:
text += ("`chunked-prefill-validation.json` records CPU comparisons with the original\n"
"graph on real text (prefill lengths across chunk boundaries and a cached\n"
diff --git a/scripts/models/bonsai/make_chunked_bundle.py b/scripts/models/bonsai/make_chunked_bundle.py
index eb502d2c7..6330adef0 100644
--- a/scripts/models/bonsai/make_chunked_bundle.py
+++ b/scripts/models/bonsai/make_chunked_bundle.py
@@ -21,6 +21,7 @@ import argparse
import json
import shutil
import tempfile
+import uuid
from pathlib import Path
from build_vision import digest, link_or_copy, read_json, seal_manifest, write_json
@@ -56,11 +57,25 @@ def make_staging(output):
def replace_output(staging, output):
- """Swap the sealed staging bundle into place; only a previous derived bundle is removed."""
+ """Swap the sealed staging bundle into place; only a previous derived bundle is removed.
+
+ The previous bundle is renamed aside first and deleted only after the new
+ one is in place, so a locked file (for example a browser still serving the
+ old bundle) cannot leave a half-deleted --output.
+ """
+ backup = None
if output.exists():
check_output(staging, output)
- shutil.rmtree(output)
- staging.rename(output)
+ backup = output.with_name(f".{output.name}.{uuid.uuid4().hex}.old")
+ output.rename(backup)
+ try:
+ staging.rename(output)
+ except BaseException:
+ if backup is not None:
+ backup.rename(output)
+ raise
+ if backup is not None:
+ shutil.rmtree(backup, ignore_errors=True)
def build_bundle(source, output, manifest, args):
@@ -94,7 +109,7 @@ def build_bundle(source, output, manifest, args):
scratch.rename(output / GRAPH)
evidence = [read_json(path) for path in args.evidence]
- card_section = model_card_section(evidence, report["output_sha256"]) # rejects failed evidence
+ card_section = model_card_section(evidence, report["output_sha256"], report["chunk_tokens"]) # rejects failed evidence
write_json(output / DERIVED_MARKER, {
"rewrite": report,
"external_data_unchanged": "onnx/decoder_model_merged_q2f16.onnx_data",
diff --git a/scripts/models/bonsai/validate_chunked_gdn.py b/scripts/models/bonsai/validate_chunked_gdn.py
index 3fbd37532..94b5fc3b6 100644
--- a/scripts/models/bonsai/validate_chunked_gdn.py
+++ b/scripts/models/bonsai/validate_chunked_gdn.py
@@ -79,6 +79,8 @@ def compare(a: dict, b: dict) -> dict:
lb = b["logits"][:, -1].astype(np.float32)
worst_state = 0.0
worst_name = ""
+ worst_fro = 0.0
+ worst_fro_name = ""
for k in a:
if k == "logits":
continue
@@ -88,6 +90,10 @@ def compare(a: dict, b: dict) -> dict:
err = float(np.abs(x - y).max()) / scale
if err > worst_state:
worst_state, worst_name = err, k
+ # Whole-tensor error: robust to a single FP16 rounding flip on an outlier.
+ fro = float(np.linalg.norm(x - y)) / max(float(np.linalg.norm(x)), 1e-6)
+ if fro > worst_fro:
+ worst_fro, worst_fro_name = fro, k
pa = np.exp(la - la.max()); pa /= pa.sum()
pb = np.exp(lb - lb.max()); pb /= pb.sum()
return {
@@ -101,6 +107,8 @@ def compare(a: dict, b: dict) -> dict:
"kl_original_to_rewritten": float((pa * (np.log(pa + 1e-30) - np.log(pb + 1e-30))).sum()),
"worst_cache_rel_err": worst_state,
"worst_cache_name": worst_name,
+ "worst_cache_rel_fro": worst_fro,
+ "worst_cache_rel_fro_name": worst_fro_name,
"finite": bool(all(np.isfinite(v.astype(np.float32)).all() for v in b.values())),
}