[codex] Add cactus CQ packaging workflow

README.md

@@ -103,6 +103,30 @@ python scripts/quantization/quantize_fixed_av.py \
   --embed-bits 4
 ```
 
+### Audio-Only Quantization Recipe
+
+Use the joint AV calibration artifacts, but leave vision and language linears
+at FP16 when you only want to quantize the audio path. The production convention
+keeps PLI at CQ2 and token embeddings at CQ4.
+
+```bash
+python scripts/quantization/quantize_fixed_av.py \
+  --model-dir "$TQH_MODEL_DIR" \
+  --calib-dir data/hessians/joint_av \
+  --out-tag audio_cq4_llm_fp16_pli2_emb4 \
+  --vision-bits 16 \
+  --audio-bits 4 \
+  --bridge-bits 4 \
+  --llm-bits 16 \
+  --pli-bits 2 \
+  --embed-bits 4
+```
+
+For a bit-width sweep, rerun with `--audio-bits 1`, `2`, `3`, and `4`, keeping
+`--bridge-bits 4 --pli-bits 2 --embed-bits 4` fixed. To measure language and
+audio interaction, set `--llm-bits 4`; to isolate the audio tower, keep
+`--llm-bits 16`.
+
 Run fixed text quantization configs:
 
 ```bash

scripts/export/export_cactus_cq.py

@@ -123,6 +123,24 @@ def fit_row_table_scale(table_fp32: torch.Tensor, token_ids: list[int]) -> torch
         "pli_bits": 2,
         "embed_bits": 4,
     },
+    "prod_v4_l2": {
+        "llm_bits": 2,
+        "vision_bits": 4,
+        "audio_bits": 4,
+        "vision_bridge_bits": 4,
+        "audio_bridge_bits": 4,
+        "pli_bits": 2,
+        "embed_bits": 4,
+    },
+    "prod_v4_l1": {
+        "llm_bits": 1,
+        "vision_bits": 4,
+        "audio_bits": 4,
+        "vision_bridge_bits": 4,
+        "audio_bridge_bits": 4,
+        "pli_bits": 2,
+        "embed_bits": 4,
+    },
     "prod_v4_l4": {
         "llm_bits": 4,
         "vision_bits": 4,