GLM-4.7_Compressed-Tensors

GLM-4.7 — Quantized (compressed-tensors for vLLM, MoE finetune)

This repository provides quantized runtime builds of
zai-org/GLM-4.7 (a Mixture-of-Experts model), repackaged for vLLM using the compressed-tensors format.

Why this quant is different (MoE-aware calibration)

• During calibration we activate all experts (not only the few picked by the router for each token).
• This captures worst-case activations across experts, yielding more robust scales and lower quant error drift at inference—especially under prompts that trigger atypical experts.
• The accompanying PR to llm-compressor adds GLM-family modeling hooks to support this MoE-aware flow (router/expert handling aligned with GLM architectures).
• Net effect: cleaner perplexity and stability, fewer edge-case artifacts, and better domain transfer when novel experts fire.

TL;DR

• Quantized with W4A16 (INT4 weights / A16 activations) for vLLM via --quantization compressed-tensors.
• Three branches by group size: W4A16_GS32, W4A16_GS64, W4A16_GS128.
• Calibration: 512 chat samples, 2048 max sequence length, dataset neuralmagic/LLM_compression_calibration (messages rendered with the model’s chat template).
• Weight-only AWQ; lm_head kept high-precision; exported with save_compressed=True.

---

Revisions & Branches

The main branch is a landing page (model card + links). Runnable artifacts live in per-quant branches.

• main — placeholder / landing page
• W4A16_GS32 — INT4 weights, group size 32 (highest fidelity; most scales)
• W4A16_GS64 — INT4, group size 64 (balanced default)
• W4A16_GS128 — INT4, group size 128 (leanest scales; fastest/lowest VRAM)

Quick links

• main: https://huggingface.co/TheHouseOfTheDude/GLM-4.7_Compressed-Tensors/tree/main
• W4A16_GS32: https://huggingface.co/TheHouseOfTheDude/GLM-4.7_Compressed-Tensors/tree/W4A16_GS32
• W4A16_GS64: https://huggingface.co/TheHouseOfTheDude/GLM-4.7_Compressed-Tensors/tree/W4A16_GS64
• W4A16_GS128: https://huggingface.co/TheHouseOfTheDude/GLM-4.7_Compressed-Tensors/tree/W4A16_GS128

---

What’s inside (per revision)

• Sharded quantized weights (*.safetensors) + index (model.safetensors.index.json)
• config.json with compressed-tensors metadata (weight_format, quantization, quantization_config, etc.)
• Tokenizer artifacts (tokenizer.json, tokenizer.model, merges/vocab as applicable)
• Optional: chat_template.jinja (inherits the finetune’s chat style)

Exact file lists may differ between branches — see Files and versions for each revision.

---

Quantization & calibration details (MoE-aware; same recipe family as recent cards)

Method / flow

• llmcompressor oneshot pipeline with an AWQModifier (weight-only).

MoE handling (GLM-4.7)

• Quantize Linear layers across all experts and shared projections.
• Router/gating Linear modules are quantized like other Linear layers.
• Expert activation during calibration: for each calibration batch, activate all experts to gather representative activation ranges across the full mixture (not just top-k). This improves scale robustness when rare experts are triggered at inference.

Targets / exclusions

• Targets: ["Linear"] (MHA/FFN and MoE expert linears).
• Ignore lm_head (kept high-precision).

Weights / grouping

• INT4 (num_bits=4, type="int", symmetric=True)
• Strategy: "group" with group size ∈ {32, 64, 128} depending on branch
• Activations are not quantized (runtime A16: BF16/FP16)

Calibration dataset & preprocessing

• Dataset: neuralmagic/LLM_compression_calibration, split train
• NUM_CALIBRATION_SAMPLES = 512 (random subset with fixed seed)
• MAX_SEQUENCE_LENGTH = 2048
• Each sample’s messages is rendered via the tokenizer’s
  apply_chat_template(..., tokenize=False), then tokenized with:
  • max_length=2048, truncation=True, padding=False, add_special_tokens=False

Compression call

• oneshot(..., max_seq_length=2048, num_calibration_samples=512, tokenizer=tokenizer) on the preprocessed dataset

Export for vLLM

• Saved with save_compressed=True so vLLM loads the compressed-tensors runtime layout directly

---

Why group size matters in AWQ (W4A16)

• Definition: Group size controls how many consecutive weights share one set of quantization scales.
• Trade-offs (accuracy ↔ throughput/VRAM):
  • GS32 (smallest groups): Most scale sets → highest fidelity (often best perplexity/task scores), but larger scale metadata, a bit more bandwidth, and slightly lower throughput.
  • GS64 (middle ground): Balanced quality and performance; strong default choice.
  • GS128 (largest groups): Fewest scale sets → leanest/faster (less bandwidth/metadata), with slightly higher quantization error; good for throughput-critical serving.
• MoE note: Smaller groups can especially help when different experts exhibit diverse activation statistics; GS32 tends to preserve expert-specific nuances best.

---

Context length

• Calibration context: up to 2048 tokens per sample (as above).
• Model context window: inherited from zai-org/GLM-4.7; quantization does not change rope/position encodings—only the numeric representation of the weights.

---

Quickstart — vLLM (compressed-tensors)

Install vLLM (recent version recommended):

pip install vllm

Serve (adjust to your hardware):

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
vllm serve TheHouseOfTheDude/GLM-4.7_Compressed-Tensors \
  --quantization compressed-tensors \
  --tensor-parallel-size 8 \
  --max-model-len 2048 \
  --gpu-memory-utilization 0.70 \
  --dtype bfloat16

Example Chat Completions:

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "TheHouseOfTheDude/GLM-4.7_Compressed-Tensors",
    "messages": [
      {"role":"system","content":"You are GLM-4.7 — helpful, precise, and safe."},
      {"role":"user","content":"Outline a plan for multi-document retrieval with MoE models."}
    ],
    "max_tokens": 512,
    "temperature": 0.7,
    "top_p": 0.95
  }'

Note: compressed-tensors is a vLLM runtime format. Loading directly with vanilla 🤗 Transformers is not supported.
For Transformers, use a compatible export (e.g., GPTQ/AWQ for Transformers) or the full-precision finetune.

---

Prompting / chat template

This package follows the finetuned parent’s chat conventions. If a chat_template.jinja is present, libraries that support apply_chat_template will automatically format messages.

Guidelines:

• Keep the system message concise (behavior, tone, safety constraints).
• Provide clear user instructions; for multi-step tasks, list steps explicitly.

---

Intended use & safety

This quantization:

• Does not change underlying behavior or content tendencies.
• Only changes weight storage for efficient inference.

Apply appropriate content filters / policies for your deployment context.

---

Lineage

• Finetuned parent: https://huggingface.co/zai-org/GLM-4.7
• This repo: Quantized child of the finetune (compressed-tensors for vLLM)

---

Hardware tips

• 100B-class MoE models benefit from multi-GPU tensor parallel; interconnect bandwidth (NVLink/IB) matters.
• Long contexts are KV-cache heavy — tune --max-model-len and batch size.
• Prefer BF16 on GPUs with native support; otherwise FP16.
• Consider CUDA Graphs if stable in your environment.

---

Changelog

• v1 (current) — Initial compressed-tensors W4A16 quantization with 512-sample / 2048-token MoE-aware calibration; branches W4A16_GS32 / W4A16_GS64 / W4A16_GS128 published; vLLM-ready packaging.