Bug: Qwen3-8B integration issue

[918822019]

incoh_rotate=true degrades W4 PPL on Qwen3-8B (opposite to paper findings on Llama-2)

Summary

When applying KronQ W4 quantization to Qwen3-8B, enabling incoh_rotate=true consistently increases PPL degradation compared to incoh_rotate=false (1.22x–1.41x vs 1.11x), which is the opposite of the expected behavior demonstrated on Llama-2-7B in the paper.

I verified that the same pipeline correctly reproduces the paper's conclusion on Llama-2-7B (1.02x with rotation vs 1.27x without), so the issue appears to be architecture-specific rather than a bug in the implementation.

Environment

• Python 3.12.9
• PyTorch 2.6.0+cu124
• CUDA 12.4
• 2× NVIDIA A100 80GB
• transformers 5.13.1
• fast-hadamard-transform 1.1.0

Model Details

Property	Value
Model	Qwen3-8B
hidden_size	4096 (power of 2 ✓)
intermediate_size	12288 (12×1024, Hadamard-compatible ✓)
num_hidden_layers	36
Architecture notes	Uses QK-Norm (RMSNorm on Q/K before attention)

Reproduction

Gradient Precomputation

# Dual-GPU parallel, layers split at 18
CUDA_VISIBLE_DEVICES=0 python -m algorithms.quantization.kronq._vendor.precompute_gradients \
  --model pretrained/Qwen3-8B \
  --output outputs/kronq_grads_qwen3_8b_wikitext \
  --calib_data datasets/wikitext_calib/data.jsonl \
  --nsamples 128 --seqlen 512 --fisher_samples 1 --batch_size 8 \
  --layer_start 0 --layer_end 18 &

CUDA_VISIBLE_DEVICES=1 python -m algorithms.quantization.kronq._vendor.precompute_gradients \
  --model pretrained/Qwen3-8B \
  --output outputs/kronq_grads_qwen3_8b_wikitext \
  --calib_data datasets/wikitext_calib/data.jsonl \
  --nsamples 128 --seqlen 512 --fisher_samples 1 --batch_size 8 \
  --layer_start 18 --layer_end 36 &
wait

PPL Evaluation

# incoh_rotate=false (best result)
python scripts/eval_kronq_ppl.py \
  --model pretrained/Qwen3-8B \
  --calib datasets/wikitext_calib/data.jsonl \
  --test data/wikitext_salesforce/wikitext-2-raw-v1/test-00000-of-00001.parquet \
  --grad_dir outputs/kronq_grads_qwen3_8b_wikitext \
  --w_bits 4 --alpha 0.1 --nsamples 64 --seqlen 512 \
  --eval_seqlen 2048 --batch_size 8

# incoh_rotate=true
python scripts/eval_kronq_ppl.py \
  --model pretrained/Qwen3-8B \
  --calib datasets/wikitext_calib/data.jsonl \
  --test data/wikitext_salesforce/wikitext-2-raw-v1/test-00000-of-00001.parquet \
  --grad_dir outputs/kronq_grads_qwen3_8b_wikitext \
  --w_bits 4 --alpha 0.25 --nsamples 64 --seqlen 512 --incoh_rotate \
  --eval_seqlen 2048 --batch_size 8

Results

Evaluation: wikitext-2 test full data, eval_seqlen=2048, 145 samples.

Config	w_bits	incoh_rotate	alpha	FP16 PPL	Quantized PPL	Degradation
A	4	false	0.1	9.72	10.83	1.11x
B	4	true	0.0	9.72	13.67	1.41x
C	4	true	0.1	9.72	12.62	1.30x
D	4	true	0.25	9.72	11.84	1.22x
E	4	true (grad v2)	0.25	9.72	12.41	1.28x

Config E uses higher-quality gradients (nsamples=128, seqlen=512, batch_size=8, dual-GPU parallel), but PPL did not improve.

Comparison with Llama-2-7B (same pipeline, same codebase)

Model	incoh_rotate=true	incoh_rotate=false	Paper conclusion reproduced?
Llama-2-7B	5.61 (1.02x)	6.99 (1.27x)	✅ Yes
Qwen3-8B	11.84 (1.22x, best)	10.83 (1.11x)	❌ No (reversed)

Analysis & Hypothesis

The key architectural difference between Llama-2 and Qwen3 is that Qwen3 uses QK-Norm (RMSNorm applied to Q and K projections before attention computation). My hypothesis:

1. The incoherent rotation (Hadamard transform) applied to weight matrices assumes that the activation distribution benefits from being "spread out" across dimensions.
2. QK-Norm already normalizes the query/key representations, potentially making the activation distribution more uniform in attention layers.
3. When BiIP's rotation is applied on top of QK-Norm's normalization, the combined effect may distort the learned representation geometry rather than improve quantization-friendliness.
4. This could explain why higher alpha (more H_G influence) partially recovers performance (Config D: 1.22x vs Config B: 1.41x at alpha=0) — the gradient covariance helps compensate, but cannot fully overcome the architectural mismatch.

Questions

1. Has KronQ been tested on Qwen-family models (especially those with QK-Norm)? Are there known compatibility considerations?
2. Would it be possible to selectively disable incoherent rotation for attention layers (where QK-Norm is active) while keeping it for MLP layers?
3. Is there a recommended approach to tune BiIP behavior for architectures with built-in normalization on Q/K?

Additional Context

• I also tested Qwen3-0.6B (hidden_size=1024, no QK-Norm issue at this scale), where incoh_rotate=true works well (1.04x vs 1.15x for false). This suggests the interaction with QK-Norm may become more pronounced at larger scales.
• Gradient quality (nsamples/seqlen) does not appear to be the bottleneck — Config E with 2x more calibration data showed no improvement.
• No NaN issues were observed on Qwen3-8B (unlike Qwen3-0.6B at alpha=0.25, which required reducing to 0.1).

[donghyunli]

Thank you for the question and the detailed experiments! We didn't implement the Qwen family, and the QK-Norm point is interesting.

First, could you set --incoh_mode full and make sure --w_clip, --w_asym, and --act_order are enabled in your command? I believe most of the PPL mismatch comes from these. Here is our reproduction of your table with the same recipe (nsamples=128, seqlen=2048):

incoh_rotate	alpha	your PPL	our PPL
false	0.1	10.83	10.627
true	0.0	13.67	10.350
true	0.1	12.62	10.375
true	0.25	11.84	10.303

Also, I found that the current gradient precomputation is not accurate for the Qwen family because of the QK-Norm: Qwen applies an RMSNorm right after q_proj/k_proj, so a standard backward hook does not capture their output gradient correctly and H_G for those two projections comes out slightly off. I fixed the precomputation code, so you can try it. With the fixed H_G, the results improve further (in our main results, alpha 0.25 has the best results):

alpha	H_G before fix	H_G after fix
0.0	10.631	10.350
0.25	10.387	10.303
0.5	10.493	10.384

---

Q2 & Q3: We ran a sublayer-group ablation on Qwen3-8B W44 (alpha 0.25, corrected H_G), applying BiIP only to selected sublayers and plain GPTAQ elsewhere.

BiIP applied to	WikiText-2 PPL
q/k only	10.183
attention (q/k/v/o)	10.192
all sublayers	10.303
none (GPTAQ)	10.650
all except q/k	10.802
MLP only	10.808

The result is the opposite of the QK-Norm-conflict intuition. On Qwen3-8B, the BiIP gain comes almost entirely from q_proj/k_proj (the QK-Norm projections), v/o are neutral, and MLP rotation is slightly harmful here. Disabling rotation on the QK-Norm layers keeps the harmful part and drops the beneficial part (10.802, worse than no rotation at all).

Interestingly, rotation on q/k alone slightly beats rotating everything on this model, which suggests per-sublayer incoherence selection could be a useful extension. We may explore this.

We made a qwen-support branch with full Qwen3 support: https://github.com/Intelligent-Computing-Lab-Panda/KronQ/tree/qwen-support. It includes the QK-Norm-aware gradient precomputation and the selective BiIP ablation (KRONQ_INCOH_GROUP=qk|no_qk|attn|mlp). See scripts/run_qwen3.sh for the exact recipe.

Please leave a comment here if you have further questions or issues.
Thank you!