Add Qwen3-30B-A3B OSCAR KV rotations

Qwen3-30B-A3B/README.md

@@ -0,0 +1,22 @@
+# Qwen3-30B-A3B OSCAR KV rotations
+
+This model needs **per-head** rotations. One rotation shared across its four KV
+heads collapses long generations (issue #16): the heads want mutually
+near-orthogonal rotations (`mean |diag(R0^T R1)| ~= 0.07`).
+
+| file | format | contents |
+|---|---|---|
+| `k_rotation_perhead.pt`, `v_rotation_perhead.pt` | `format_version: 2` | `rotation: [4, 128, 128]` per layer, 48 layers -- **use these** |
+| `k_rotation_qqt_r_h_pbr.pt`, `v_rotation_sst_r_h_pbr.pt` | `format_version: 1` | one shared `[128, 128]` per layer -- for comparison only |
+
+3 seeds, TP=2, 32K generation:
+
+| Benchmark | BF16 | INT2 shared | INT2 per-head |
+|---|:---:|:---:|:---:|
+| GPQA | 60.8 +- 3.2 | 43.9 +- 2.8 | **58.6 +- 0.9** |
+| HumanEval | 90.9 +- 2.1 | 25.8 +- 2.5 | **89.0 +- 1.2** |
+| AIME 2025 | 70.0 +- 6.7 | 18.9 +- 10.7 | **65.6 +- 1.9** |
+| MATH500 | 93.6 +- 0.9 | 60.8 +- 2.2 | **92.1 +- 0.4** |
+
+Serving per-head needs a runtime that reads `format_version: 2`; the pool logs
+`[per-head: 4 kv heads]` at startup when it is active.

Qwen3-30B-A3B/k_rotation_perhead.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf819e3d5d437e44e0245c2da065482d27d642d2cd2a7672428f91e375d4898a
+size 12597631

Qwen3-30B-A3B/k_rotation_qqt_r_h_pbr.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec433986663dffad5f6646ab686f0b74f5dbd6c3555eb32d3bd8fd432d584603
+size 3198717

Qwen3-30B-A3B/v_rotation_perhead.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d72de2cfc02bfffd5562f88f38cda75dd83e680bc4e6ab40995183d5e89371f
+size 12597631

Qwen3-30B-A3B/v_rotation_sst_r_h_pbr.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:adaf48e0b85952e53baeeaf9d54f1e37f11101ba773f401ebc942f19700e7617
+size 3198717