fix: lower seqlen to 512 for short calibration examples

Training examples are ~500-1500 tokens. seqlen=2048 causes 'no data has been cached' error. Also remove deprecated format param.
2026-06-01 04:27:24 +02:00
parent 465e74f49e
commit 2200120133
1 changed files with 1 additions and 2 deletions
@@ -46,10 +46,9 @@ rounder = AutoRound(
    dataset=calib,
    bits=4,
    group_size=128,
-    seqlen=2048,
+    seqlen=512,
    nsamples=min(128, len(calib)),
    iters=200,
-    format="auto_round",
 )

 print("Running AutoRound INT4 quantization...")