vh/esh-pfi-infrastructure
1
0
Fork 0
Code Issues 7 Pull Requests Actions Packages Projects Releases Wiki Activity

fix(gen-seat): hash bf16 tensors via uint8 reinterpret, not numpy

Browse Source 
numpy has no bfloat16, so .numpy().tobytes() raised
'TypeError: Got unsupported ScalarType BFloat16' on real checkpoints.
Flatten then view(torch.uint8) before hashing.

Result on the heresy candidate: VERDICT IDENTICAL -- all 15 mtp.* tensors
byte-identical to the incumbent's verbatim base graft, the head already
measured at 47.7% acceptance in production. The ~56 GB bf16 acceptance gate
is redundant, so no second seat comes down.
This commit is contained in:
vh
2026-08-17 16:29:45 -07:00
parent 254c588921
commit 2c3602869f
1 changed files with 10 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
services/gen-seat-mixed-quant/compare_mtp_head.py
+10 -2
View File
@@ -74,8 +74,16 @@ def load_mtp(model_dir: Path) -> dict:
def digest(tensor) -> str: def digest(tensor) -> str:
"""SHA-256 over the raw tensor bytes. Dtype-sensitive by design — a head """SHA-256 over the raw tensor bytes. Dtype-sensitive by design — a head
stored at a different precision is not the same head for our purposes.""" stored at a different precision is not the same head for our purposes.
return hashlib.sha256(tensor.contiguous().view(-1).numpy().tobytes()).hexdigest()
Goes through a uint8 reinterpret rather than `.numpy()`: numpy has no
bfloat16, and these checkpoints are bf16, so the direct route raises
`TypeError: Got unsupported ScalarType BFloat16`.
"""
import torch
flat = tensor.contiguous().flatten()
return hashlib.sha256(flat.view(torch.uint8).numpy().tobytes()).hexdigest()
def main() -> int: def main() -> int: