MrSphay/odysseus
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: hwfit native quant labels miss the cost maps and over-estimate VRAM (#1690)

Browse Source
This commit is contained in:
Afonso Coutinho
2026-06-03 06:22:42 +01:00
committed by GitHub
parent 77b92020c0
commit 3d00c85636
2 changed files with 48 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
services/hwfit/fit.py
View File

@@ -280,10 +280,14 @@ def _native_quant(model):
return "FP8"
if "gptq" in text:
m = re.search(r"(?:gptq|int|w)(?:[-_]?)(\d{1,2})(?:bit)?", text)
return f"GPTQ-{m.group(1)}bit" if m else "GPTQ"
# Canonical catalog label is "GPTQ-Int4"/"GPTQ-Int8" (see models.py
# QUANT_BPP / QUANT_QUALITY_PENALTY keys); "GPTQ-4bit" misses both
# maps, so BPP and the quality penalty silently fall to defaults.
return f"GPTQ-Int{m.group(1)}" if m else "GPTQ-Int4"
if "awq" in text:
m = re.search(r"(?:awq|int|w)(?:[-_]?)(\d{1,2})(?:bit)?", text)
return f"AWQ-{m.group(1)}bit" if m else "AWQ"
# Catalog keys are "AWQ-4bit"/"AWQ-8bit"; bare "AWQ" misses the maps.
return f"AWQ-{m.group(1)}bit" if m else "AWQ-4bit"
if "mlx" in text:
m = re.search(r"mlx[-_]?(\d{1,2})bit", text)
return f"mlx-{m.group(1)}bit" if m else native_quant