From f72b18fda0d84d26046791bb9b37e4ccd441b27c Mon Sep 17 00:00:00 2001
From: Linxiushen <1787979356@qq.com>
Date: Sun, 20 Sep 2026 18:03:13 +0800
Subject: [PATCH] fix: quant_analyse crashes on per-channel/group weight scales

PTQ.save() gates a scale-outlier report behind quantization.quant_analyse and
tests each scale with `if act_scales_data > 1.5:` / `if weight_scales_data > 1.5:`.
That only works for 0-dim scalars. Per-channel and group-wise quantization keep
one scale per channel/group, so the tensor holds more than one element and
Python calls Tensor.__bool__(), which raises:

    RuntimeError: Boolean value of Tensor with more than one value is ambiguous

This is reachable from the shipped configs: configs/qwen3/ptq/int8_dynamic/*.yaml
all set weight: "per-channel", and AbsMaxChannelWiseWeightObserver computes
max(dim=quant_axis()) without keepdim, i.e. a 1-D per-channel tensor.

Use torch.any() so both 0-dim and multi-element scales work, and report
.max().item() so the message stays short instead of printing a whole
per-channel tensor.

per-tensor behaviour is unchanged (still warns above 1.5, silent below).
---
 angelslim/compressor/quant/ptq.py | 15 +++++++++++----
 1 file changed, 11 insertions(+), 4 deletions(-)

diff --git a/angelslim/compressor/quant/ptq.py b/angelslim/compressor/quant/ptq.py
index d4377d17..ee698e60 100644
--- a/angelslim/compressor/quant/ptq.py
+++ b/angelslim/compressor/quant/ptq.py
@@ -194,20 +194,27 @@ def save(self, save_path: str):
             and self.quant_model.quant_config.quant_analyse
         ):
             # scale analyse
+            # Scales are not always 0-dim: per-channel / group-wise quantization
+            # keeps one scale per channel or group, so these tensors usually hold
+            # more than one element (the int8_dynamic configs default to
+            # `weight: "per-channel"`). `if tensor > 1.5` then raises
+            # "Boolean value of Tensor with more than one value is ambiguous".
+            # torch.any() handles both shapes, and .max() keeps the message short
+            # instead of printing a whole per-channel tensor.
             for k in self.quant_model.act_scales_dict.keys():
                 act_scales_data = self.quant_model.act_scales_dict[k].data
-                if act_scales_data > 1.5:
+                if torch.any(act_scales_data > 1.5):
                     print_info(
                         f"[AngelSlim Warning] Act_scales {k}: "
-                        f"The weight is too high:{act_scales_data}. "
+                        f"The weight is too high:{act_scales_data.max().item()}. "
                         f"It is recommended to clip it to 1.5 "
                     )
             for k in self.quant_model.weight_scales_dict.keys():
                 weight_scales_data = self.quant_model.weight_scales_dict[k].data
-                if weight_scales_data > 1.5:
+                if torch.any(weight_scales_data > 1.5):
                     print_info(
                         f"[AngelSlim Warning] Weight_scales {k}: "
-                        f"The weight is too high:{weight_scales_data}. "
+                        f"The weight is too high:{weight_scales_data.max().item()}. "
                         f"It is recommended to clip it to 1.5 "
                     )