fix: quant_analyse crashes on per-channel/group weight scales

[Linxiushen]

Problem

PTQ.save() gates a scale-outlier report behind quantization.quant_analyse and tests each scale like this (angelslim/compressor/quant/ptq.py, lines 199 and 207):

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:          # <-- only valid for a 0-dim scalar

if tensor > 1.5 works only when the scale is a 0-dim scalar. Per-channel and group-wise quantization keep one scale per channel/group, so the tensor holds more than one element, Python falls back to Tensor.__bool__(), and it 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" (checked qwen3-0_6b, qwen3-1_7b, qwen3-32b)
• AbsMaxChannelWiseWeightObserver._cal_min_max computes new_inp.abs().max(dim=self.quant_axis()) without keepdim, i.e. a 1-D per-channel tensor
• quant_analyse is wired into both the int8 and fp8 branches (quant/core/config.py), so it is meant to be a general switch, not per-tensor-only

So enabling quant_analyse on the default int8_dynamic recipe crashes at save time, after calibration has already run.

Reproduction

The block below is extracted verbatim from ptq.py at main (ee8ddb2) and executed against a per-channel scale tensor shaped exactly like the observer's output:

# current main
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:
        print_info(...)

# fed: {"model.layers.0.mlp.down_proj.weight": torch.tensor([0.1, 2.0, 0.3, 0.05])}
RuntimeError: Boolean value of Tensor with more than one value is ambiguous

With the patch applied, the same block reports normally:

[AngelSlim Warning] Weight_scales model.layers.0.mlp.down_proj.weight: The weight is too high:2.0. ...

Minimal standalone check of the root cause:

>>> torch.tensor(2.0) > 1.5 and True          # 0-dim: fine
>>> bool(torch.tensor([0.1, 2.0]) > 1.5)
RuntimeError: Boolean value of Tensor with more than one value is ambiguous

Fix

Use torch.any(), which handles 0-dim and multi-element tensors alike, and print .max().item() so the warning stays readable instead of dumping a whole per-channel tensor.

Behaviour across scale shapes

scale	before	after
per-tensor, above threshold (0-dim)	warns	warns (max=2.0)
per-tensor, below threshold (0-dim)	silent	silent
per-channel, some above (4 elem)	RuntimeError	warns (max=5.0)
per-channel, none above (4 elem)	RuntimeError	silent
per-group 2-D (2×3)	RuntimeError	warns (max=9.9)
single-element 1-D	warns	warns (max=2.0)

per-tensor behaviour is unchanged in both directions — that was the main thing I wanted to be sure of.

Notes

• 11 insertions / 4 deletions, one file
• torch is already imported at the top of ptq.py (line 19), no new dependency
• I could not run a full end-to-end quantization to confirm (no GPU / no triton wheel on this machine), so the verification above executes the real source block directly rather than driving PTQ.save() through a real calibration run. Happy to adjust if you'd like the warning formatted differently, e.g. also printing the offending channel index.

---

Investigated and fixed with AI assistance (Claude); I reviewed the change and the verification above myself.