HawkBearPig/GLM-5.3-Int4-Int8Mix-AWQ-g64

GLM-5.3 Int4/Int8 mixed-precision AWQ, group 64

A mixed-precision quantization of GLM-5.3 (754B-parameter MoE, model type glm_moe_dsa) made from the original BF16 release (zai-org/GLM-5.3-BF16, not the derived FP8 checkpoint). Built for the DGPP inference engine on four NVIDIA DGX Sparks (TP=4), where the whole model has to fit in ~100 GiB per rank with room left for context.

Total size	398 GiB (≈ 99.8 GiB per rank at TP=4)
Routed experts, layers 3–77 (mlp.experts.*.{gate,up,down}_proj)	int4 symmetric, group 64, bf16 scales, AWQ
Attention (q_a, q_b, kv_a_proj_with_mqa, kv_b, o_proj) and shared experts, layers 3–77	int8 symmetric, group 64, bf16 scales
Dense layers 0–2, DSA indexers, MoE routers, norms, embeddings, lm_head, the whole MTP block (layer 78)	bf16, byte-exact from the source
Format	compressed-tensors pack-quantized (weight_packed int32, weight_scale bf16, weight_shape), declared in config.json

Method

• AWQ (activation-aware weight quantization) via llm-compressor 0.13.0, applied one decoder layer at a time by a layer-wise driver that is bit-exact against llm-compressor's single-pass sequential pipeline (verified on a reduced model, both group sizes). Activations propagate through the smoothed, unquantized layer, as in the reference implementation.
• Calibration: 512 sequences × 2048 tokens (1.05M tokens) drawn with a fixed seed from a mixed corpus of prose, code, math and conversational / tool-use transcripts. Each routed expert is calibrated on the tokens actually routed to it. A second, independent 512-row draw reproduced the layer output error to within 0.1 %, so the calibration size is not a limiting factor.
• Fold correction (important). llm-compressor's AWQ mapping for glm_moe_dsa folds the smoothing scale of post_attention_layernorm into the norm but does not rescale the MoE router (mlp.gate), and likewise leaves the DSA indexer's wk / weights_proj / wq_b uncompensated in the 18 layers that own an indexer. As saved by llm-compressor those modules receive x / s, and the per-layer error is 5–10× worse than plain round-to-nearest. Every MoE shard here has the affected weight columns rescaled by the recovered scale (original_norm / smoothed_norm), which restores the layer to slightly better than round-to-nearest. Corrected shards carry the safetensors metadata key fold_corrected. The defect and fix are described in the issue draft shipped with this card's tooling.

Measured error (per layer)

Relative RMS error against the bf16 layer on 32 × 2048 real, held-out tokens. "Contribution" is the error of what the layer adds (output minus input); "residual" is the error of the full residual stream leaving the layer, which is the quantity that propagates.

layer	variant	contribution	residual
3	this checkpoint (AWQ int4 g64, corrected)	2.38 %	0.53 %
3	RTN int4 g64 (no AWQ)	2.40 %	0.54 %
3	AWQ as llm-compressor saved it (uncorrected, g128)	12.5 %	2.80 %
3	fp8 e4m3 g128 on every projection, for reference	2.46 %	0.55 %
6 (owns an indexer)	this checkpoint	3.27 %	1.50 %
6	RTN int4 g64	3.44 %	1.58 %
6	AWQ uncorrected	30.5 %	14.0 %
21	this checkpoint	10.95 %	0.58 %
21	RTN int4 g64	11.09 %	0.59 %
42 (owns an indexer)	this checkpoint	9.85 %	2.00 %
42	RTN int4 g64	9.89 %	2.01 %
63	this checkpoint	12.50 %	2.34 %
63	RTN int4 g64	12.50 %	2.34 %

The contribution metric grows with depth because each layer's addition shrinks relative to the residual stream; the residual error rises from 0.5 % (layer 3) to 2.3 % (layer 63). On this model the int4/int8 mix lands at the error of an all-fp8 recipe at half the bytes. AWQ's advantage over round-to-nearest is 0.9 / 4.8 / 1.2 / 0.4 / 0.02 % of the error at layers 3 / 6 / 21 / 42 / 63: real in the shallow indexer layer, negligible at depth. Per layer, this checkpoint and the round-to-nearest companion are equivalent.

No end-to-end evaluation yet (perplexity, benchmarks): the numbers above are per-layer. Treat this as a carefully validated but not yet benchmarked checkpoint.

A companion checkpoint with the same recipe but plain round-to-nearest instead of AWQ (no smoothing, no fold correction needed) is at HawkBearPig/GLM-5.3-Int4-Int8Mix-RTN-g64.

Loading

Standard compressed-tensors layout: model.safetensors.index.json, per-layer shards layer-NNN.safetensors, passthrough.safetensors (embeddings, norms, lm_head, MTP block), config.json with quantization_config, tokenizer files. Any loader that understands glm_moe_dsa and compressed-tensors pack-quantized W4A16/W8A16 should read it; only DGPP has been exercised. vLLM / transformers loading is untested.

Tooling: llm-compressor 0.13.0, compressed-tensors 0.18.0, transformers 5.14.1, torch 2.13.0 (CUDA 13), one B200.

License: MIT, inherited from GLM-5.3.