Publish PrismQuant paper, concise model card, and loader manifest

README.md

@@ -12,18 +12,23 @@ tags:
 - gptq
 - w4a4kv4
 - custom-code
-- research
+- arxiv:2609.32429
 ---
 
-# Qwen3 Base — PrismQuant
+# PrismQuant-Qwen3-Base
 
-Released checkpoints for [PrismQuant](https://github.com/ForeverBlue816/PrismQuant): quantizer-aware rotations for low-bit language models.
+[Paper](https://arxiv.org/abs/2609.32429) · [Code](https://github.com/ForeverBlue816/PrismQuant) · [Loading guide](https://github.com/ForeverBlue816/PrismQuant/blob/main/docs/models.md)
 
-**Base models:** [Qwen/Qwen3-0.6B-Base](https://huggingface.co/Qwen/Qwen3-0.6B-Base), [Qwen/Qwen3-1.7B-Base](https://huggingface.co/Qwen/Qwen3-1.7B-Base), [Qwen/Qwen3-4B-Base](https://huggingface.co/Qwen/Qwen3-4B-Base), [Qwen/Qwen3-8B-Base](https://huggingface.co/Qwen/Qwen3-8B-Base).
+Official checkpoints for **PrismQuant: Optimal Null-Space Rotations for Grouped Quantizers**. PrismQuant aligns dominant activation directions with the constant directions of asymmetric quantization groups.
 
-**arXiv:** coming soon.
+## Models
 
-Model weights retain [Apache-2.0](LICENSE).
+| Model key | Base model | Quantization |
+| --- | --- | --- |
+| `qwen3_0.6b_base` | [Qwen/Qwen3-0.6B-Base](https://huggingface.co/Qwen/Qwen3-0.6B-Base) | W4A4KV4 |
+| `qwen3_1.7b_base` | [Qwen/Qwen3-1.7B-Base](https://huggingface.co/Qwen/Qwen3-1.7B-Base) | W4A4KV4 |
+| `qwen3_4b_base` | [Qwen/Qwen3-4B-Base](https://huggingface.co/Qwen/Qwen3-4B-Base) | W4A4KV4 |
+| `qwen3_8b_base` | [Qwen/Qwen3-8B-Base](https://huggingface.co/Qwen/Qwen3-8B-Base) | W4A4KV4 |
 
 ## Quick start
 
@@ -31,7 +36,6 @@ Model weights retain [Apache-2.0](LICENSE).
 git clone --depth 1 https://github.com/ForeverBlue816/PrismQuant.git
 cd PrismQuant
 pip install -e .
-prismquant models
 ```
 
 ```python
@@ -41,39 +45,16 @@ loaded = load_model("qwen3_0.6b_base")
 print(loaded.generate("The key idea behind quantization is", max_new_tokens=64))
 ```
 
-The loader selectively downloads one checkpoint and its required rotation factors, pins its revision, and obtains the matching base model/tokenizer. Defaults use PrismQuant `k=max`, seed 0. Use `prismquant models` to select another exact checkpoint name. See the [loading guide](https://github.com/ForeverBlue816/PrismQuant/blob/main/docs/models.md).
+The loader selects the default PrismQuant checkpoint, downloads its required files at a pinned revision, and obtains the matching base model and tokenizer. These are base models for text completion.
 
-## Format and limitations
+## Checkpoint format
 
-These are **dequantized GPTQ INT4 weights stored in floating-point tensors**, not packed INT4 weights or standalone Transformers checkpoints. Do not pass this repository directly to `AutoModelForCausalLM.from_pretrained`. The PrismQuant loader restores unquantized state from the base model, loads the saved decoder linears, and installs the online rotations and activation/KV quantizers.
+The GPTQ INT4 values are stored as **dequantized floating-point tensors**. The reference runtime applies activation and KV quantization while retaining floating-point storage. Use the PrismQuant loader; this repository is not a standalone Transformers checkpoint or a packed INT4 serving model.
 
-The reference runtime retains floating-point weight and KV storage. It uses the original bf16 containers for Llama 3B/8B and fp32 for Qwen3/70B. W4A4KV4 names the simulated numerical quantization, not this artifact's disk size or serving memory. Dedicated packed-kernel deployment benchmarks are documented separately in the code repository.
+Reference compute: **float32**. Budget for the base-model state, activations and KV cache as well as the downloaded tensors.
 
-These are base-model completions rather than chat-tuned assistants. The original evaluation protocols, seeds, caveats and negative findings remain in the [research archive](https://github.com/ForeverBlue816/PrismQuant/tree/research-archive-2026-09-25). Quantization does not remove upstream model limitations.
+`config.json` declares the PrismQuant artifact format and default models. `checkpoints/` contains decoder weights; `rotations/` contains the factors required by the loader.
 
-## Contents
+## License
 
-- `checkpoints/<model>/<checkpoint>/layer_XX.pt`: the folded, GPTQ-quantized decoder linear tensors.
-- `DONE.json` and `gptq_audit.csv`: quantization settings and provenance.
-- `rotations/<model>/...`: the calibrated factors used by the runtime.
-
-The `nar` strings in filenames are the original internal name of PrismQuant and are retained for compatibility. Existing weight files are unchanged by this public-release update. The two smaller Llama repositories additionally include the previously omitted `k=8` down-projection factors.
-
-## Available checkpoints
-
-Download sizes below cover one variant and its required factors, excluding the separately downloaded base model.
-
-| Model key | Checkpoint | Weight protocol | Download (GB) |
-| --- | --- | --- | --- |
-| `qwen3_0.6b_base` | `gptq_hadamard_seed0_g128_asym` | g128_asym | 1.76 |
-| `qwen3_0.6b_base` | `gptq_nar_k8_seed0_g128_asym` | g128_asym | 1.77 |
-| `qwen3_0.6b_base` | `gptq_nar_kmax_seed0_g128_asym` | g128_asym | 1.77 |
-| `qwen3_1.7b_base` | `gptq_hadamard_seed0_g128_asym` | g128_asym | 5.64 |
-| `qwen3_1.7b_base` | `gptq_nar_k8_seed0_g128_asym` | g128_asym | 5.65 |
-| `qwen3_1.7b_base` | `gptq_nar_kmax_seed0_g128_asym` | g128_asym | 5.67 |
-| `qwen3_4b_base` | `gptq_hadamard_seed0_g128_asym` | g128_asym | 14.53 |
-| `qwen3_4b_base` | `gptq_nar_k8_seed0_g128_asym` | g128_asym | 14.55 |
-| `qwen3_4b_base` | `gptq_nar_kmax_seed0_g128_asym` | g128_asym | 14.65 |
-| `qwen3_8b_base` | `gptq_hadamard_seed0_g128_asym` | g128_asym | 27.78 |
-| `qwen3_8b_base` | `gptq_nar_k8_seed0_g128_asym` | g128_asym | 27.80 |
-| `qwen3_8b_base` | `gptq_nar_kmax_seed0_g128_asym` | g128_asym | 27.96 |
+Qwen-derived weights retain [Apache-2.0](LICENSE).

config.json

@@ -0,0 +1,34 @@
+{
+  "format": "prismquant-reference-v1",
+  "paper": "https://arxiv.org/abs/2609.32429",
+  "models": {
+    "qwen3_0.6b_base": {
+      "base_model": "Qwen/Qwen3-0.6B-Base",
+      "default_checkpoint": "gptq_nar_kmax_seed0_g128_asym",
+      "architecture": "qwen3",
+      "quantization": "W4A4KV4",
+      "storage_dtype": "float32"
+    },
+    "qwen3_1.7b_base": {
+      "base_model": "Qwen/Qwen3-1.7B-Base",
+      "default_checkpoint": "gptq_nar_kmax_seed0_g128_asym",
+      "architecture": "qwen3",
+      "quantization": "W4A4KV4",
+      "storage_dtype": "float32"
+    },
+    "qwen3_4b_base": {
+      "base_model": "Qwen/Qwen3-4B-Base",
+      "default_checkpoint": "gptq_nar_kmax_seed0_g128_asym",
+      "architecture": "qwen3",
+      "quantization": "W4A4KV4",
+      "storage_dtype": "float32"
+    },
+    "qwen3_8b_base": {
+      "base_model": "Qwen/Qwen3-8B-Base",
+      "default_checkpoint": "gptq_nar_kmax_seed0_g128_asym",
+      "architecture": "qwen3",
+      "quantization": "W4A4KV4",
+      "storage_dtype": "float32"
+    }
+  }
+}