add the Monarch block-count sweep: monarch_5 / monarch_10 / monarch_20

.gitattributes

@@ -46,3 +46,6 @@ blockdiag_8/g_best filter=lfs diff=lfs merge=lfs -text
 blockdiag_full/g_best filter=lfs diff=lfs merge=lfs -text
 blockdiag_fc/g_best filter=lfs diff=lfs merge=lfs -text
 wide_blockdiag/g_best filter=lfs diff=lfs merge=lfs -text
+monarch_10/g_best filter=lfs diff=lfs merge=lfs -text
+monarch_20/g_best filter=lfs diff=lfs merge=lfs -text
+monarch_5/g_best filter=lfs diff=lfs merge=lfs -text

README.md

@@ -70,6 +70,26 @@ does survive the fix — but it had never actually been tested before it.
 | `monarch_fc`   | 2.38 M |     2.843 |     2.831 |        +0.012 |
 | `monarch_full` | 1.10 M |     2.838 |     2.846 |        −0.009 |
 
+#### Monarch block-count sweep (`nblocks` 5 / 8 / 10 / 20)
+
+Same architecture as `monarch_8` — hidden 400, fc 600, both FCs and both GRU
+projections Monarch — with **only `nblocks` varying**, so parameters move
+without any other change. RTF is int8 on CPU; the three new runs were measured
+together on an idle box.
+
+| run          | nblocks | params | FP32 PESQ | int8 PESQ | Δ (FP32→int8) | int8 RTF |
+| ------------ | ------: | -----: | --------: | --------: | ------------: | -------: |
+| `monarch_5`  |       5 | 0.88 M |     2.852 |     2.858 |        −0.007 |    0.017 |
+| `monarch_8`  |       8 | 0.55 M | **2.861** |     2.856 |        +0.005 |    0.027 |
+| `monarch_10` |      10 | 0.44 M |     2.849 |     2.842 |        +0.007 |    0.014 |
+| `monarch_20` |      20 | 0.23 M |     2.849 |     2.854 |        −0.005 |    0.013 |
+
+**4× parameters, 0.012 PESQ spread** — the 0.23 M `monarch_20` matches the
+0.88 M `monarch_5`, and int8 stays loss-free (|Δ| ≤ 0.007) even at the smallest
+size. `monarch_20` is the cheapest model in the family. The `monarch_8` RTF was
+measured in an earlier session under different load, so it is not comparable to
+the other three.
+
 ### Block-diagonal, dense, butterfly
 
 | run                 | params | FP32 PESQ | int8 PESQ | Δ (FP32→int8) |
@@ -94,7 +114,9 @@ does survive the fix — but it had never actually been tested before it.
   NSNet2 predicts a magnitude mask and reuses the noisy phase, which caps PESQ
   regardless of how expressive the mask predictor is. **The dense model was
   already over-parameterized** — which is exactly why aggressive structuring is
-  nearly free. For deployment, take the smallest (`blockdiag_8` / `monarch_8`).
+  nearly free. For deployment, take the smallest (`blockdiag_8` / `monarch_20`).
+  The block-count sweep extends this: 4× parameters across `nblocks` 5→20 moves
+  PESQ by 0.012, within run-to-run noise.
 - **Genuine Monarch beats block-diagonal, but marginally** (+0.011…+0.038 FP32 at
   matched `nblocks`) and it costs parameters — its second factor makes it larger.
   Consistent with the saturation above.
@@ -113,7 +135,8 @@ static int8 ONNX, and the exact `config.json` it was trained with.
 baseline/          blockdiag_8/     monarch_8/       butterfly_fc/
 blockdiag_fc/      blockdiag_full/  monarch_fc/      butterfly_full/
 wide_blockdiag/    monarch_full/    wide_monarch/    butterfly_ortho/
-                                                     butterfly_2blocks/
+                   monarch_5/       monarch_10/      butterfly_2blocks/
+                   monarch_20/
 
   each: {g_best, g_best_fp32.onnx, g_best.onnx, config.json}
 ```

monarch_10/config.json

@@ -0,0 +1,37 @@
+{
+    "num_gpus": 0,
+    "batch_size": 256,
+    "learning_rate": 0.003,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.99,
+    "seed": 1234,
+    "hidden_dim": 400,
+    "fc_hidden_dim": 600,
+    "num_gru_layers": 2,
+    "compress_factor": 0.3,
+    "linear": {
+        "kind": "monarch",
+        "nblocks": 10
+    },
+    "gru": {
+        "kind": "triton_monarch",
+        "nblocks": 10,
+        "struct_input": true
+    },
+    "sampling_rate": 16000,
+    "segment_size": 32000,
+    "n_fft": 512,
+    "hop_size": 256,
+    "win_size": 512,
+    "num_workers": 8,
+    "quant": {
+        "enabled": false,
+        "n_calib_utts": 200
+    },
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1
+    }
+}

monarch_10/g_best

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b358c5c5de3b3ac8de436ff0da76d3f99a5653253069ff29c24fd7b0017f308
+size 1833745

monarch_10/g_best.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d63d2b56ad3b8f18d8a213e890ddee31f3710adb19eb8e6134efb427daaba560
+size 566366

monarch_10/g_best_fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f9a194bc3374b8972b06f28df110417312b3bfa2c72e66e52315bece772b7d3
+size 1805283

monarch_20/config.json

@@ -0,0 +1,37 @@
+{
+    "num_gpus": 0,
+    "batch_size": 256,
+    "learning_rate": 0.003,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.99,
+    "seed": 1234,
+    "hidden_dim": 400,
+    "fc_hidden_dim": 600,
+    "num_gru_layers": 2,
+    "compress_factor": 0.3,
+    "linear": {
+        "kind": "monarch",
+        "nblocks": 20
+    },
+    "gru": {
+        "kind": "triton_monarch",
+        "nblocks": 20,
+        "struct_input": true
+    },
+    "sampling_rate": 16000,
+    "segment_size": 32000,
+    "n_fft": 512,
+    "hop_size": 256,
+    "win_size": 512,
+    "num_workers": 8,
+    "quant": {
+        "enabled": false,
+        "n_calib_utts": 200
+    },
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1
+    }
+}

monarch_20/g_best

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c10a754e8361cd88810143b7a811d6cf8328e4120386e14d6f57e9d6d8d661d3
+size 960273

monarch_20/g_best.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f8c73b0de34db72a6ce4b820059c9ab0bfd3ebf1dcf465d1a349255cb4fe57d2
+size 345174

monarch_20/g_best_fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ca0f2802c8032fa043bbd19fcbbb59d7c96b4e4579b373296ee024172a6bb0f
+size 931761

monarch_5/config.json

@@ -0,0 +1,37 @@
+{
+    "num_gpus": 0,
+    "batch_size": 256,
+    "learning_rate": 0.003,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.99,
+    "seed": 1234,
+    "hidden_dim": 400,
+    "fc_hidden_dim": 600,
+    "num_gru_layers": 2,
+    "compress_factor": 0.3,
+    "linear": {
+        "kind": "monarch",
+        "nblocks": 5
+    },
+    "gru": {
+        "kind": "triton_monarch",
+        "nblocks": 5,
+        "struct_input": true
+    },
+    "sampling_rate": 16000,
+    "segment_size": 32000,
+    "n_fft": 512,
+    "hop_size": 256,
+    "win_size": 512,
+    "num_workers": 8,
+    "quant": {
+        "enabled": false,
+        "n_calib_utts": 200
+    },
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1
+    }
+}

monarch_5/g_best

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69496ae315a21000fd3e471f4e2fda1bce16aff23c7ae2a339eb0ee5cf6eac1c
+size 3580753

monarch_5/g_best.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f04861d3e15f3d87bcaf884de8cafa1e103f289d7bc905d4785d3395e2941e7
+size 1008147

monarch_5/g_best_fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a8b383e6d25a18f71787d74cd8eee264052ff01a0b2f5c52dff899cff175816
+size 3552324