rename monarch_*->blockdiag_* (they were single block-diagonal factors), publish genuine 2-factor Monarch runs, re-quantize int8 with Einsum weight fix

.gitattributes

@@ -42,3 +42,7 @@ monarch_8/g_best filter=lfs diff=lfs merge=lfs -text
 monarch_fc/g_best filter=lfs diff=lfs merge=lfs -text
 monarch_full/g_best filter=lfs diff=lfs merge=lfs -text
 wide_monarch/g_best filter=lfs diff=lfs merge=lfs -text
+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

README.md

@@ -7,6 +7,7 @@ tags:
   - speech-denoising
   - nsnet2
   - butterfly
+  - block-diagonal
   - monarch
   - structured-matrices
   - onnx
@@ -22,96 +23,115 @@ metrics:
 
 Best-PESQ checkpoints from the
 [eco8-neaixt](https://github.com/LarocheC/eco8-neaixt) compression sweep:
-NSNet2 speech enhancement with the FC and GRU layers swappable between
-dense, [Butterfly](https://arxiv.org/abs/1903.05895), and
-[Monarch](https://arxiv.org/abs/2204.00595) factorizations.
+NSNet2 speech enhancement with the FC and GRU layers swappable between dense,
+[Butterfly](https://arxiv.org/abs/1903.05895), **block-diagonal**, and genuine
+two-factor [Monarch](https://arxiv.org/abs/2204.00595) factorizations.
 
 Trained on
-[VoiceBank-DEMAND-16k](https://huggingface.co/datasets/JacobLinCool/VoiceBank-DEMAND-16k)
-for 200 epochs at batch 256, n\_fft 512, on a single GTX 1080 Ti.
-
-Each variant ships in three formats:
-- `g_best` — PyTorch checkpoint (training-time weights)
-- `g_best_fp32.onnx` — streaming-shape FP32 ONNX (frame-by-frame, opset 17)
-- `g_best.onnx` — static int8 ONNX (QDQ format, per-channel weight quant,
-  MinMax calibration on 200 VBD-train utterances)
-
-## Results
-
-PESQ measured on the full VBD test split (824 utterances). RTF (real-time
-factor) is for the int8 ONNX session under onnxruntime CPU; lower is faster.
-
-| run                 | params  | FP32 PESQ |  int8 PESQ |  Δ (FP32→int8) | int8 RTF |
-| ------------------- | ------: | --------: | ---------: | -------------: | -------: |
-| `wide_monarch`      |  2.36 M | **2.864** |      2.842 |        +0.021  |    0.166 |
-| `baseline`          |  2.78 M |     2.845 |      2.833 |        +0.012  |    0.452 |
-| `monarch_8`         |  0.36 M |     2.832 |      2.826 |        +0.006  | **0.025** |
-| `monarch_full`      |  0.70 M |     2.827 |  **2.848** |        −0.021  |    0.039 |
-| `monarch_fc`        |  2.14 M |     2.805 |      2.789 |        +0.016  |    0.448 |
-| `butterfly_2blocks` |  0.36 M |     2.805 |      2.202 |        +0.602  |    0.441 |
-| `butterfly_fc`      |  1.99 M |     2.799 |      2.494 |        +0.306  |    0.522 |
-| `butterfly_ortho`   |  0.19 M |     2.780 |      2.577 |        +0.203  |    0.232 |
-| `butterfly_full`    |  0.19 M |     2.772 |      2.128 |        +0.644  |    0.230 |
+[VoiceBank-DEMAND-16k](https://huggingface.co/datasets/JacobLinCool/VoiceBank-DEMAND-16k),
+n\_fft 512, batch 256. PESQ on the full 824-utterance test split.
+
+## ⚠️ Two corrections you should read before quoting these numbers
+
+**1. Naming: the old `monarch_*` runs were NOT Monarch.** They are a *single
+block-diagonal factor* (one block-diagonal matrix per projection, **zero
+cross-block mixing**). A genuine [Monarch](https://arxiv.org/abs/2204.00595) is a
+**two-factor** construction (block-diagonal × permutation × block-diagonal) with
+full cross-channel mixing. The old runs have therefore been **renamed to
+`blockdiag_*`**, and the `monarch_*` names now hold *genuinely* Monarch models:
+
+| old name (was mislabeled) | now                | `monarch_*` today        |
+| ------------------------- | ------------------ | ------------------------ |
+| `monarch_8`               | → `blockdiag_8`    | genuine 2-factor Monarch |
+| `monarch_full`            | → `blockdiag_full` | genuine 2-factor Monarch |
+| `monarch_fc`              | → `blockdiag_fc`   | genuine 2-factor Monarch |
+| `wide_monarch`            | → `wide_blockdiag` | genuine 2-factor Monarch |
+
+If you previously pinned `monarch_8`, you now get a **different (genuinely
+Monarch) model** — the block-diagonal one you had is at `blockdiag_8`.
+
+**2. int8: the previously published int8 ONNX never quantized the structured
+weights.** The structure-preserving export lowers each block-diagonal / Monarch
+matmul to an `Einsum`, and onnxruntime ships no QDQ handler for `Einsum` — so
+`quantize_static` skipped those nodes entirely. Only activations and the residual
+dense MatMuls were int8; the **dominant weights stayed FP32**. Every int8 file
+here has been **re-quantized with the structured weights genuinely int8** (fixed
+in `nsnet2/qdq_einsum_quantizer.py` upstream). The "int8 is loss-free" property
+does survive the fix — but it had never actually been tested before it.
+
+## Results (int8 = genuinely quantized weights)
+
+### Genuine two-factor Monarch
+
+| run            | params | FP32 PESQ | int8 PESQ | Δ (FP32→int8) |
+| -------------- | -----: | --------: | --------: | ------------: |
+| `wide_monarch` | 3.64 M | **2.881** | **2.884** |        −0.003 |
+| `monarch_8`    | 0.55 M |     2.861 |     2.856 |        +0.005 |
+| `monarch_fc`   | 2.38 M |     2.843 |     2.831 |        +0.012 |
+| `monarch_full` | 1.10 M |     2.838 |     2.846 |        −0.009 |
+
+### Block-diagonal, dense, butterfly
+
+| run                 | params | FP32 PESQ | int8 PESQ | Δ (FP32→int8) |
+| ------------------- | -----: | --------: | --------: | ------------: |
+| `wide_blockdiag`    | 2.36 M |     2.864 |     2.847 |        +0.016 |
+| `baseline` (dense)  | 2.78 M |     2.845 |     2.833 |        +0.012 |
+| `blockdiag_8`       | 0.36 M |     2.832 |     2.825 |        +0.007 |
+| `blockdiag_full`    | 0.70 M |     2.827 |     2.843 |        −0.016 |
+| `blockdiag_fc`      | 2.14 M |     2.805 |     2.787 |        +0.018 |
+| `butterfly_2blocks` | 0.36 M |     2.805 |     2.202 |        +0.602 |
+| `butterfly_fc`      | 1.99 M |     2.799 |     2.494 |        +0.306 |
+| `butterfly_ortho`   | 0.19 M |     2.780 |     2.577 |        +0.203 |
+| `butterfly_full`    | 0.19 M |     2.772 |     2.128 |        +0.644 |
 
 ## Key findings
 
-- **Monarch variants are essentially loss-free under int8** (|Δ| ≤ 0.021
-  across the board). A single `Einsum` per FC plus per-channel weight
-  quantization is genuinely friendly to int8 calibration.
-- **`wide_monarch` is the best deployment target** for quality: highest FP32
-  PESQ (2.864) with near-zero int8 loss (2.842). For speed-constrained
-  deployment, **`monarch_full` and `monarch_8`** trade ~0.02 PESQ for an
-  RTF of 0.04 / 0.025 — over 10× faster than the dense baseline.
-- **Butterfly with random init degrades catastrophically under int8**
-  (Δ up to 0.64 PESQ on `butterfly_full`). Int8 deployment with butterfly
-  factorizations should use `init=ortho`: `butterfly_ortho` loses 0.20
-  PESQ to int8 versus 0.64 for `butterfly_full` (same architecture, same
-  training data, only the init differs).
-- **Longer training makes randn-init butterfly *worse* on int8.** The same
-  `butterfly_full` config saw its int8 gap grow from 0.36 PESQ at 50
-  epochs → 0.64 PESQ at 200 epochs as twiddle factors drifted further
-  from orthogonality. Ortho-init butterfly does not show this regression.
+- **Quality saturates — this model class is architecture-bound, not
+  capacity-bound.** Across three structure families and ~10× parameters, every
+  configuration lands in a **2.83–2.88** band, non-monotonically: the 0.55 M
+  `monarch_8` (2.861) beats both the 1.10 M `monarch_full` and the 2.38 M
+  `monarch_fc`. Going 7× from `monarch_8` to `wide_monarch` buys +0.020 PESQ.
+  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`).
+- **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.
+- **Block-diagonal and Monarch quantize loss-free** (|Δ| ≤ 0.018 and ≤ 0.012),
+  *with the weights genuinely quantized*.
+- **Butterfly with randn init degrades catastrophically under int8** (Δ up to
+  0.644). Use `init=ortho`: `butterfly_ortho` loses 0.203 to int8 vs 0.644 for
+  `butterfly_full` — same architecture, same data, only the init differs.
 
 ## Layout
 
-Each subdirectory is one run, containing the saved generator (`g_best`),
-the streaming FP32 ONNX, the static int8 ONNX, and the exact `config.json`
-the run was trained with.
+One subdirectory per run: the generator (`g_best`), the streaming FP32 ONNX, the
+static int8 ONNX, and the exact `config.json` it was trained with.
 
 ```
-baseline/{g_best,g_best_fp32.onnx,g_best.onnx,config.json}
-monarch_fc/{g_best,g_best_fp32.onnx,g_best.onnx,config.json}
-butterfly_fc/{g_best,g_best_fp32.onnx,g_best.onnx,config.json}
-monarch_full/{g_best,g_best_fp32.onnx,g_best.onnx,config.json}
-butterfly_full/{g_best,g_best_fp32.onnx,g_best.onnx,config.json}
-monarch_8/{g_best,g_best_fp32.onnx,g_best.onnx,config.json}
-butterfly_ortho/{g_best,g_best_fp32.onnx,g_best.onnx,config.json}
-butterfly_2blocks/{g_best,g_best_fp32.onnx,g_best.onnx,config.json}
-wide_monarch/{g_best,g_best_fp32.onnx,g_best.onnx,config.json}
+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/
+
+  each: {g_best, g_best_fp32.onnx, g_best.onnx, config.json}
 ```
 
 ## Loading
 
-### PyTorch checkpoint
-
-Clone the repo first (model classes live there):
-
 ```bash
-git clone https://github.com/LarocheC/eco8-neaixt
-cd eco8-neaixt
-uv sync
+git clone https://github.com/LarocheC/eco8-neaixt && cd eco8-neaixt && uv sync
 ```
 
-Then:
-
 ```python
 import json, torch
 from huggingface_hub import hf_hub_download
-from env import AttrDict
-from models.model import NSNet2
+from common.env import AttrDict
+from nsnet2.model import NSNet2
 
 REPO = "claroche1/sparse-nsnet2-checkpoints"
-RUN  = "wide_monarch"  # or any name from the table
+RUN  = "monarch_8"   # or any run from the tables
 
 cfg  = json.load(open(hf_hub_download(REPO, f"{RUN}/config.json")))
 ckpt = torch.load(hf_hub_download(REPO, f"{RUN}/g_best"),
@@ -121,53 +141,43 @@ model = NSNet2(AttrDict(cfg)).cuda().eval()
 model.load_state_dict(ckpt["generator"])
 ```
 
+> **Note for the `monarch_*` runs:** their GRU was trained through gru-qat's fused
+> Monarch Triton kernel (`"gru": {"kind": "triton_monarch"}`), so their
+> `state_dict` uses gru-qat module names. Loading them requires
+> **`gru-qat >= 0.5.0`** and **`torch-structured >= 1.3.0`** (both pulled in by
+> `uv sync`). The `blockdiag_*` / `butterfly_*` / `baseline` runs use the native
+> path and have no such requirement.
+
 ### ONNX (FP32 or int8)
 
-The ONNX models are streaming-shape: a single frame `(B, n_freq)` plus the
-GRU state `(num_layers, B, hidden)` per session call, threaded across
-frames. The end-to-end pipeline (RMS-norm → STFT → frame loop → iSTFT)
-is in `inference_onnx.py` in the source repo.
+Streaming-shape: one frame `(B, n_freq)` plus GRU state `(num_layers, B, hidden)`
+per session call, threaded across frames. End-to-end pipeline in
+`nsnet2/inference_onnx.py`.
 
 ```python
 import onnxruntime as ort
 from huggingface_hub import hf_hub_download
-
-REPO = "claroche1/sparse-nsnet2-checkpoints"
-RUN  = "wide_monarch"
-
-# FP32:
-fp32_path = hf_hub_download(REPO, f"{RUN}/g_best_fp32.onnx")
-fp32_sess = ort.InferenceSession(fp32_path, providers=["CPUExecutionProvider"])
-
-# int8 (deployment):
-int8_path = hf_hub_download(REPO, f"{RUN}/g_best.onnx")
-int8_sess = ort.InferenceSession(int8_path, providers=["CPUExecutionProvider"])
+REPO, RUN = "claroche1/sparse-nsnet2-checkpoints", "monarch_8"
+sess = ort.InferenceSession(hf_hub_download(REPO, f"{RUN}/g_best.onnx"),
+                            providers=["CPUExecutionProvider"])   # int8
 ```
 
-End-to-end inference example with PESQ measurement is in `inference_onnx.py`
-in the source repo.
-
 ## Citations
 
 ```bibtex
 @inproceedings{braun2021nsnet2,
     title={Towards efficient models for real-time deep noise suppression},
     author={Braun, Sebastian and Tashev, Ivan},
-    booktitle={ICASSP},
-    year={2021}
+    booktitle={ICASSP}, year={2021}
 }
-
 @inproceedings{dao2019butterfly,
     title={Learning fast algorithms for linear transforms using butterfly factorizations},
     author={Dao, Tri and Gu, Albert and Eichhorn, Matthew and Rudra, Atri and R{\'e}, Christopher},
-    booktitle={ICML},
-    year={2019}
+    booktitle={ICML}, year={2019}
 }
-
 @inproceedings{dao2022monarch,
     title={Monarch: Expressive structured matrices for efficient and accurate training},
     author={Dao, Tri and Chen, Beidi and Sohoni, Nimit S and Desai, Arjun and Poli, Michael and Grogan, Jessica and Liu, Alexander and Rao, Aniruddh and Rudra, Atri and R{\'e}, Christopher},
-    booktitle={ICML},
-    year={2022}
+    booktitle={ICML}, year={2022}
 }
 ```

blockdiag_8/config.json

@@ -0,0 +1,36 @@
+{
+    "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": "blockdiag",
+        "nblocks": 8
+    },
+    "gru": {
+        "kind": "blockdiag",
+        "nblocks": 8
+    },
+    "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
+    }
+}

blockdiag_8/g_best

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c7416829373b77c1a0fd9ef7a33681131abe33cda0bdf95e8ddc585610a7995b
+size 1423570

blockdiag_8/g_best.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b179824dcf6eb6e04bb1d685b58ada46be7af398cc8a31cc16edadbd83ce1e86
+size 432971

blockdiag_8/g_best_fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ceaa0e96895fcfab277c8b0409fe0d9299a6f60f52bbd8fa6cc5e454b6e1ce73
+size 1442161

blockdiag_fc/config.json

@@ -0,0 +1,35 @@
+{
+    "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": "blockdiag",
+        "nblocks": 4
+    },
+    "gru": {
+        "kind": "gru"
+    },
+    "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
+    }
+}

blockdiag_fc/g_best

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cde0e25bebea3d5d713afaaf10f5e52eabaeb5ef552de45abc1cb216957e1848
+size 8570351

blockdiag_fc/g_best.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:264bb5059e5359cd15b6f22fd1990fbfc537df6040be62ba9d8829ef5540b4b2
+size 2252655

blockdiag_fc/g_best_fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f105f2a4d8d2974d9c7da1264c58404e311ed5538fd7ea14c688cdb9610eb5f6
+size 8578184

blockdiag_full/config.json

@@ -0,0 +1,36 @@
+{
+    "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": "blockdiag",
+        "nblocks": 4
+    },
+    "gru": {
+        "kind": "blockdiag",
+        "nblocks": 4
+    },
+    "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
+    }
+}

blockdiag_full/g_best

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c9a34fe8ba46ad1c41ac474010eca2a8bbaadf30d334901af81146b93cc95ba7
+size 2811538

blockdiag_full/g_best.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a666da906fccb95d99b452cc996a4ea446712aa4ccf73eda94408488832e866e
+size 784123

blockdiag_full/g_best_fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c49a7c8648612dc6db1acadbd2077db15f6f7197bea47d4c175d0b3cfc2b6894
+size 2830166

monarch_8/config.json

@@ -15,8 +15,9 @@
         "nblocks": 8
     },
     "gru": {
-        "kind": "monarch",
-        "nblocks": 8
+        "kind": "triton_monarch",
+        "nblocks": 8,
+        "struct_input": true
     },
     "sampling_rate": 16000,
     "segment_size": 32000,
@@ -33,4 +34,4 @@
         "dist_url": "tcp://localhost:54321",
         "world_size": 1
     }
-}
+}

monarch_8/g_best

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c7416829373b77c1a0fd9ef7a33681131abe33cda0bdf95e8ddc585610a7995b
-size 1423570
+oid sha256:bb927b19edc8fc94da1b4cedc52dc1a75b56ab13272b9cd3d0101b5f1ab28353
+size 2273489

monarch_8/g_best.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:368f9a6af401f42000294c45d5b827dd275d9fb2c486f2e687943d3f23451251
-size 1465674
+oid sha256:37174cc619ecbb73e2526b4ec853d8fc97c6dea0969bf0c1d1358ac5d5409b61
+size 677653

monarch_8/g_best_fp32.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0776227333ef5075c967846daae9e3ecbdc8f924dc0905002a6b5f00703eec49
-size 1442161
+oid sha256:94d37f1c9ba921dde8b06ef8f4a20cbcce37ac5ab71f9a94c68f058563e51b3f
+size 2245092

monarch_fc/config.json

@@ -15,7 +15,7 @@
         "nblocks": 4
     },
     "gru": {
-        "kind": "gru"
+        "kind": "triton"
     },
     "sampling_rate": 16000,
     "segment_size": 32000,
@@ -32,4 +32,4 @@
         "dist_url": "tcp://localhost:54321",
         "world_size": 1
     }
-}
+}

monarch_fc/g_best

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cde0e25bebea3d5d713afaaf10f5e52eabaeb5ef552de45abc1cb216957e1848
-size 8570351
+oid sha256:3be49fc25099bd925764cf2d7449417f96ca885ae0e2f90d92c3c2eb6eb7627e
+size 9555373

monarch_fc/g_best.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:febb340361f7f9c015f45ebdb031dee822a6e652eed48f18382c140cac878122
-size 2891983
+oid sha256:0f03d3c0cb744e0d4b2f2ce55f813e21ca6b0be59446d21a548e472433b58ef2
+size 2510827

monarch_fc/g_best_fp32.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1bda161d596109d21b794656776329f4986cdd78be7362edafbb2b57534cae16
-size 8578184
+oid sha256:82d9cac18e0cdffa56caa906b62d0daa7e4fe852b589b5af89712849a6fd81e8
+size 9537912

monarch_full/config.json

@@ -15,8 +15,9 @@
         "nblocks": 4
     },
     "gru": {
-        "kind": "monarch",
-        "nblocks": 4
+        "kind": "triton_monarch",
+        "nblocks": 4,
+        "struct_input": true
     },
     "sampling_rate": 16000,
     "segment_size": 32000,
@@ -33,4 +34,4 @@
         "dist_url": "tcp://localhost:54321",
         "world_size": 1
     }
-}
+}

monarch_full/g_best

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c9a34fe8ba46ad1c41ac474010eca2a8bbaadf30d334901af81146b93cc95ba7
-size 2811538
+oid sha256:d595a5f69cf0c40f92cae8e23a545fcc70c0c0cad1305a4dbff02cbe2e09a1f1
+size 4454353

monarch_full/g_best.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:640d277c5843df8c4e1b0518e20c29a7f377847a436b62f02ae0ff43e4b6cd04
-size 2853656
+oid sha256:f06d4607c1ffa9bd9d056cfef88c06afba7e09bdb06c95beb3ea165bb402d0e2
+size 1229091

monarch_full/g_best_fp32.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:accb5c6dfc79dde3f7772ed90b67ce9b9dc3fcb06b26de9784d4f0d100099b64
-size 2830166
+oid sha256:7df98544d085c0c99157bcf707704fc8542b07f737dcd6c6a333cf76e51bbd8e
+size 4425852

wide_blockdiag/config.json

@@ -0,0 +1,36 @@
+{
+    "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": 768,
+    "fc_hidden_dim": 1024,
+    "num_gru_layers": 2,
+    "compress_factor": 0.3,
+    "linear": {
+        "kind": "blockdiag",
+        "nblocks": 4
+    },
+    "gru": {
+        "kind": "blockdiag",
+        "nblocks": 4
+    },
+    "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
+    }
+}

wide_blockdiag/g_best

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e9ad0dda8f30258d8a9dfa42fd900c4e418bc97bdc2a2ed9a6aae709ab6542d7
+size 9432786

wide_blockdiag/g_best.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b0bf6ce3a0cd7650f647dcff733cbd47aa9c8645f1bf04cc82088b5ada975cb
+size 2446614

wide_blockdiag/g_best_fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:993a4691187c94e50c5e2de7a52a29113529cc0a487b8d98d989212594a7c9fa
+size 9451516

wide_monarch/config.json

@@ -15,8 +15,9 @@
         "nblocks": 4
     },
     "gru": {
-        "kind": "monarch",
-        "nblocks": 4
+        "kind": "triton_monarch",
+        "nblocks": 4,
+        "struct_input": true
     },
     "sampling_rate": 16000,
     "segment_size": 32000,
@@ -33,4 +34,4 @@
         "dist_url": "tcp://localhost:54321",
         "world_size": 1
     }
-}
+}

wide_monarch/g_best

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e9ad0dda8f30258d8a9dfa42fd900c4e418bc97bdc2a2ed9a6aae709ab6542d7
-size 9432786
+oid sha256:7cd2e11239d38449de74f26abb802772a467de331939ed2dfa8a3248edcef490
+size 14602065

wide_monarch/g_best.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cc5ac41831080f05d188ca695b9b57a77bf2ee3092008be5fd9ce64f5ecfeb58
-size 9458215
+oid sha256:86d3c6cd8df5c77ed14eee425d6d25c716728450e8ea6a5c1e2b16cca6c9014c
+size 3776607

wide_monarch/g_best_fp32.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1af01d6864dbb59f459cd25dc77b68befb28d7963f1164dffcb42b300167fa3b
-size 9451516
+oid sha256:656b4c78c4c032e4913fb04604cb56492cf15cb934bd5a92e5182349f4854937
+size 14573504