Upload folder using huggingface_hub

flash_attn/artifacts/benchmark/Attention Benchmark.csv

@@ -1,7 +1,7 @@
 seq_len,torch_cudnn,torch_cudnn_compile_d,torch_cudnn_compile_ma,torch_flash,torch_flash_compile_d,torch_flash_compile_ma,hf_flash_attn,hf_flash_attn3
-4224.000000,3.807456,3.789232,4.191984,3.974816,3.953792,4.322096,3.403408,3.328416
-4352.000000,4.078480,4.072352,4.420736,4.400480,4.390000,4.738144,3.833424,3.755664
-4416.000000,4.139680,4.134800,4.490464,4.451040,4.443680,4.795104,3.890112,3.860992
-4480.000000,4.202048,4.195216,4.561248,4.524608,4.519520,4.877056,3.948816,3.866704
-4544.000000,4.434992,4.427040,4.788000,4.582336,4.571872,4.945728,4.015280,3.982320
-4608.000000,4.499456,4.490816,4.874464,4.669152,4.663648,5.035232,4.071872,3.983520
+4224.000000,3.801472,3.790064,4.182320,3.968000,3.957824,4.311152,3.398160,3.330400
+4352.000000,4.082944,4.082912,4.413488,4.400000,4.391936,4.738048,3.837424,3.758208
+4416.000000,4.142624,4.135648,4.484160,4.452304,4.446096,4.792480,3.892064,3.864128
+4480.000000,4.206144,4.198752,4.551808,4.530752,4.522944,4.873760,3.949344,3.870224
+4544.000000,4.438320,4.433104,4.787584,4.584160,4.576640,4.934304,4.008960,3.974672
+4608.000000,4.502432,4.495456,4.871872,4.660192,4.651040,5.029792,4.065616,3.984160

flash_attn/benchmark.html

@@ -3722,7 +3722,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <span onclick="toggleOutput('nv')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-nv" style="cursor: default; opacity: 0.3;">▶ uv-logs</span>
 </span> | 
-Cell: nv | 0.70s
+Cell: nv | 0.67s
  | <button class="run-btn" onclick="runCell('nv')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('nv')">Copy</button>
 <a href="cells/nv.py" target="_blank" class="raw-btn">Raw</a>
@@ -3745,7 +3745,7 @@ Cell: nv | 0.70s
 </div>
 </div>
 <div id="output-nv" class="cell-output">
-<div class="cell-stdout">Fri Sep 26 02:23:10 2025       
+<div class="cell-stdout">Fri Sep 26 03:53:23 2025       
 +-----------------------------------------------------------------------------------------+
 | NVIDIA-SMI 570.172.08             Driver Version: 570.172.08     CUDA Version: 12.8     |
 |-----------------------------------------+------------------------+----------------------+
@@ -3754,19 +3754,19 @@ Cell: nv | 0.70s
 |                                         |                        |               MIG M. |
 |=========================================+========================+======================|
 |   0  NVIDIA A10G                    On  |   00000000:00:1B.0 Off |                    0 |
-|  0%   31C    P8             26W /  300W |       0MiB /  23028MiB |      0%      Default |
+|  0%   38C    P0             51W /  300W |       0MiB /  23028MiB |      0%      Default |
 |                                         |                        |                  N/A |
 +-----------------------------------------+------------------------+----------------------+
 |   1  NVIDIA A10G                    On  |   00000000:00:1C.0 Off |                    0 |
-|  0%   30C    P8             26W /  300W |       0MiB /  23028MiB |      0%      Default |
+|  0%   31C    P8             24W /  300W |       0MiB /  23028MiB |      0%      Default |
 |                                         |                        |                  N/A |
 +-----------------------------------------+------------------------+----------------------+
 |   2  NVIDIA A10G                    On  |   00000000:00:1D.0 Off |                    0 |
-|  0%   31C    P8             27W /  300W |       0MiB /  23028MiB |      0%      Default |
+|  0%   31C    P8             25W /  300W |       0MiB /  23028MiB |      0%      Default |
 |                                         |                        |                  N/A |
 +-----------------------------------------+------------------------+----------------------+
 |   3  NVIDIA A10G                    On  |   00000000:00:1E.0 Off |                    0 |
-|  0%   30C    P8             27W /  300W |       0MiB /  23028MiB |      0%      Default |
+|  0%   31C    P8             25W /  300W |       0MiB /  23028MiB |      0%      Default |
 |                                         |                        |                  N/A |
 +-----------------------------------------+------------------------+----------------------+
 
@@ -3782,19 +3782,19 @@ Cell: nv | 0.70s
 </div>
 </div>
 
-<div class="cell cell-failed" id="cell-benchmark">
+<div class="cell" id="cell-benchmark">
 <div class="cell-header">
 <span class="collapse-indicators">
 <span onclick="toggleCode('benchmark')" style="cursor: pointer;">▼ code</span> 
 <span onclick="toggleOutput('benchmark')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-benchmark" onclick="toggleUvLogsFromHeader('benchmark')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: benchmark | 77.48s | FAILED
+Cell: benchmark | 75.46s
  | <button class="run-btn" onclick="runCell('benchmark')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('benchmark')">Copy</button>
 <a href="cells/benchmark.py" target="_blank" class="raw-btn">Raw</a>
 </div>
-<div id="code-benchmark" class="cell-code" data-lines="345">
+<div id="code-benchmark" class="cell-code" data-lines="343">
 <div class="highlight-with-lines">
 <div class="line-numbers" id="lines-benchmark">
 <a class="line-number" data-cell="benchmark" data-line="1" href="#cell-benchmark" onclick="event.preventDefault(); selectCellLine('benchmark', 1, true);">1</a>
@@ -4140,8 +4140,6 @@ Cell: benchmark | 77.48s | FAILED
 <a class="line-number" data-cell="benchmark" data-line="341" href="#cell-benchmark" onclick="event.preventDefault(); selectCellLine('benchmark', 341, true);">341</a>
 <a class="line-number" data-cell="benchmark" data-line="342" href="#cell-benchmark" onclick="event.preventDefault(); selectCellLine('benchmark', 342, true);">342</a>
 <a class="line-number" data-cell="benchmark" data-line="343" href="#cell-benchmark" onclick="event.preventDefault(); selectCellLine('benchmark', 343, true);">343</a>
-<a class="line-number" data-cell="benchmark" data-line="344" href="#cell-benchmark" onclick="event.preventDefault(); selectCellLine('benchmark', 344, true);">344</a>
-<a class="line-number" data-cell="benchmark" data-line="345" href="#cell-benchmark" onclick="event.preventDefault(); selectCellLine('benchmark', 345, true);">345</a>
 </div>
 <div class="code-wrap">
 <div class="highlight"><pre><span></span><span class="c1"># /// script</span>
@@ -4487,8 +4485,6 @@ Cell: benchmark | 77.48s | FAILED
 <span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">inference_mode</span><span class="p">():</span>
     <span class="n">correctness</span><span class="p">()</span>
     <span class="n">fig</span> <span class="o">=</span> <span class="n">benchmark_fn</span><span class="o">.</span><span class="n">run</span><span class="p">(</span><span class="n">print_data</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span> <span class="n">save_path</span><span class="o">=</span><span class="n">output_dir</span><span class="o">.</span><span class="n">as_posix</span><span class="p">())</span>
-    <span class="n">fig</span><span class="o">.</span><span class="n">savefig</span><span class="p">(</span><span class="n">output_dir</span> <span class="o">/</span> <span class="s2">&quot;attention_benchmark.png&quot;</span><span class="p">,</span> <span class="n">dpi</span><span class="o">=</span><span class="mi">300</span><span class="p">,</span> <span class="n">bbox_inches</span><span class="o">=</span><span class="s2">&quot;tight&quot;</span><span class="p">)</span>
-    <span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">&quot;Benchmark plot saved to: </span><span class="si">{</span><span class="n">output_dir</span><span class="w"> </span><span class="o">/</span><span class="w"> </span><span class="s1">&#39;attention_benchmark.png&#39;</span><span class="si">}</span><span class="s2">&quot;</span><span class="p">)</span>
 </pre></div>
 
 <div class="code-line-highlight" id="line-highlight-benchmark"></div>
@@ -4504,105 +4500,105 @@ xFormers not found.
 
 
 ===== Testing shape: (1, 4224, 24, 128) =====
-torch_cudnn                   : absmax=0.000994, mae=0.000075, mse=0.000000
-torch_cudnn_compile_d         : absmax=0.000994, mae=0.000075, mse=0.000000
-torch_cudnn_compile_ma        : absmax=0.000994, mae=0.000075, mse=0.000000
-torch_flash                   : absmax=0.000994, mae=0.000075, mse=0.000000
-torch_flash_compile_d         : absmax=0.000994, mae=0.000075, mse=0.000000
-torch_flash_compile_ma        : absmax=0.000994, mae=0.000075, mse=0.000000
-hf_flash_attn                 : absmax=0.000994, mae=0.000075, mse=0.000000
-hf_flash_attn3                : absmax=0.000994, mae=0.000075, mse=0.000000
+torch_cudnn                   : absmax=0.001524, mae=0.000075, mse=0.000000
+torch_cudnn_compile_d         : absmax=0.001524, mae=0.000075, mse=0.000000
+torch_cudnn_compile_ma        : absmax=0.001524, mae=0.000075, mse=0.000000
+torch_flash                   : absmax=0.001524, mae=0.000075, mse=0.000000
+torch_flash_compile_d         : absmax=0.001524, mae=0.000075, mse=0.000000
+torch_flash_compile_ma        : absmax=0.001524, mae=0.000075, mse=0.000000
+hf_flash_attn                 : absmax=0.001524, mae=0.000075, mse=0.000000
+hf_flash_attn3                : absmax=0.001524, mae=0.000075, mse=0.000000
 
 
 ===== Testing shape: (1, 4352, 24, 128) =====
-torch_cudnn                   : absmax=0.001718, mae=0.000073, mse=0.000000
-torch_cudnn_compile_d         : absmax=0.001718, mae=0.000073, mse=0.000000
-torch_cudnn_compile_ma        : absmax=0.001718, mae=0.000073, mse=0.000000
-torch_flash                   : absmax=0.001718, mae=0.000073, mse=0.000000
-torch_flash_compile_d         : absmax=0.001718, mae=0.000073, mse=0.000000
-torch_flash_compile_ma        : absmax=0.001718, mae=0.000073, mse=0.000000
-hf_flash_attn                 : absmax=0.001718, mae=0.000073, mse=0.000000
-hf_flash_attn3                : absmax=0.001718, mae=0.000073, mse=0.000000
+torch_cudnn                   : absmax=0.001335, mae=0.000074, mse=0.000000
+torch_cudnn_compile_d         : absmax=0.001335, mae=0.000074, mse=0.000000
+torch_cudnn_compile_ma        : absmax=0.001335, mae=0.000074, mse=0.000000
+torch_flash                   : absmax=0.001321, mae=0.000074, mse=0.000000
+torch_flash_compile_d         : absmax=0.001321, mae=0.000074, mse=0.000000
+torch_flash_compile_ma        : absmax=0.001321, mae=0.000074, mse=0.000000
+hf_flash_attn                 : absmax=0.001321, mae=0.000074, mse=0.000000
+hf_flash_attn3                : absmax=0.001321, mae=0.000074, mse=0.000000
 
 
 ===== Testing shape: (1, 4416, 24, 128) =====
-torch_cudnn                   : absmax=0.001273, mae=0.000073, mse=0.000000
-torch_cudnn_compile_d         : absmax=0.001273, mae=0.000073, mse=0.000000
-torch_cudnn_compile_ma        : absmax=0.001273, mae=0.000073, mse=0.000000
-torch_flash                   : absmax=0.001440, mae=0.000073, mse=0.000000
-torch_flash_compile_d         : absmax=0.001440, mae=0.000073, mse=0.000000
-torch_flash_compile_ma        : absmax=0.001440, mae=0.000073, mse=0.000000
-hf_flash_attn                 : absmax=0.001440, mae=0.000073, mse=0.000000
-hf_flash_attn3                : absmax=0.001440, mae=0.000073, mse=0.000000
+torch_cudnn                   : absmax=0.000897, mae=0.000073, mse=0.000000
+torch_cudnn_compile_d         : absmax=0.000897, mae=0.000073, mse=0.000000
+torch_cudnn_compile_ma        : absmax=0.000897, mae=0.000073, mse=0.000000
+torch_flash                   : absmax=0.000897, mae=0.000073, mse=0.000000
+torch_flash_compile_d         : absmax=0.000897, mae=0.000073, mse=0.000000
+torch_flash_compile_ma        : absmax=0.000897, mae=0.000073, mse=0.000000
+hf_flash_attn                 : absmax=0.000897, mae=0.000073, mse=0.000000
+hf_flash_attn3                : absmax=0.000897, mae=0.000073, mse=0.000000
 
 
 ===== Testing shape: (1, 4480, 24, 128) =====
-torch_cudnn                   : absmax=0.001284, mae=0.000073, mse=0.000000
-torch_cudnn_compile_d         : absmax=0.001284, mae=0.000073, mse=0.000000
-torch_cudnn_compile_ma        : absmax=0.001284, mae=0.000073, mse=0.000000
-torch_flash                   : absmax=0.001284, mae=0.000073, mse=0.000000
-torch_flash_compile_d         : absmax=0.001284, mae=0.000073, mse=0.000000
-torch_flash_compile_ma        : absmax=0.001284, mae=0.000073, mse=0.000000
-hf_flash_attn                 : absmax=0.001284, mae=0.000073, mse=0.000000
-hf_flash_attn3                : absmax=0.001284, mae=0.000073, mse=0.000000
+torch_cudnn                   : absmax=0.001691, mae=0.000073, mse=0.000000
+torch_cudnn_compile_d         : absmax=0.001691, mae=0.000073, mse=0.000000
+torch_cudnn_compile_ma        : absmax=0.001691, mae=0.000073, mse=0.000000
+torch_flash                   : absmax=0.001691, mae=0.000073, mse=0.000000
+torch_flash_compile_d         : absmax=0.001691, mae=0.000073, mse=0.000000
+torch_flash_compile_ma        : absmax=0.001691, mae=0.000073, mse=0.000000
+hf_flash_attn                 : absmax=0.001691, mae=0.000073, mse=0.000000
+hf_flash_attn3                : absmax=0.001691, mae=0.000073, mse=0.000000
 
 
 ===== Testing shape: (1, 4544, 24, 128) =====
-torch_cudnn                   : absmax=0.000976, mae=0.000072, mse=0.000000
-torch_cudnn_compile_d         : absmax=0.000976, mae=0.000072, mse=0.000000
-torch_cudnn_compile_ma        : absmax=0.000976, mae=0.000072, mse=0.000000
-torch_flash                   : absmax=0.000834, mae=0.000072, mse=0.000000
-torch_flash_compile_d         : absmax=0.000834, mae=0.000072, mse=0.000000
-torch_flash_compile_ma        : absmax=0.000834, mae=0.000072, mse=0.000000
-hf_flash_attn                 : absmax=0.000834, mae=0.000072, mse=0.000000
-hf_flash_attn3                : absmax=0.000815, mae=0.000072, mse=0.000000
+torch_cudnn                   : absmax=0.001201, mae=0.000072, mse=0.000000
+torch_cudnn_compile_d         : absmax=0.001201, mae=0.000072, mse=0.000000
+torch_cudnn_compile_ma        : absmax=0.001201, mae=0.000072, mse=0.000000
+torch_flash                   : absmax=0.001201, mae=0.000072, mse=0.000000
+torch_flash_compile_d         : absmax=0.001201, mae=0.000072, mse=0.000000
+torch_flash_compile_ma        : absmax=0.001201, mae=0.000072, mse=0.000000
+hf_flash_attn                 : absmax=0.001201, mae=0.000072, mse=0.000000
+hf_flash_attn3                : absmax=0.001201, mae=0.000072, mse=0.000000
 
 
 ===== Testing shape: (1, 4608, 24, 128) =====
-torch_cudnn                   : absmax=0.000926, mae=0.000072, mse=0.000000
-torch_cudnn_compile_d         : absmax=0.000926, mae=0.000072, mse=0.000000
-torch_cudnn_compile_ma        : absmax=0.000926, mae=0.000072, mse=0.000000
-torch_flash                   : absmax=0.000926, mae=0.000072, mse=0.000000
-torch_flash_compile_d         : absmax=0.000926, mae=0.000072, mse=0.000000
-torch_flash_compile_ma        : absmax=0.000926, mae=0.000072, mse=0.000000
-hf_flash_attn                 : absmax=0.000926, mae=0.000072, mse=0.000000
-hf_flash_attn3                : absmax=0.000926, mae=0.000072, mse=0.000000
+torch_cudnn                   : absmax=0.001150, mae=0.000071, mse=0.000000
+torch_cudnn_compile_d         : absmax=0.001150, mae=0.000071, mse=0.000000
+torch_cudnn_compile_ma        : absmax=0.001150, mae=0.000071, mse=0.000000
+torch_flash                   : absmax=0.001150, mae=0.000071, mse=0.000000
+torch_flash_compile_d         : absmax=0.001150, mae=0.000071, mse=0.000000
+torch_flash_compile_ma        : absmax=0.001150, mae=0.000071, mse=0.000000
+hf_flash_attn                 : absmax=0.001150, mae=0.000071, mse=0.000000
+hf_flash_attn3                : absmax=0.001150, mae=0.000071, mse=0.000000
 Attention Benchmark:
    seq_len  torch_cudnn  torch_cudnn_compile_d  torch_cudnn_compile_ma  torch_flash  torch_flash_compile_d  torch_flash_compile_ma  hf_flash_attn  hf_flash_attn3
-0   4224.0     3.807456               3.789232                4.191984     3.974816               3.953792                4.322096       3.403408        3.328416
-1   4352.0     4.078480               4.072352                4.420736     4.400480               4.390000                4.738144       3.833424        3.755664
-2   4416.0     4.139680               4.134800                4.490464     4.451040               4.443680                4.795104       3.890112        3.860992
-3   4480.0     4.202048               4.195216                4.561248     4.524608               4.519520                4.877056       3.948816        3.866704
-4   4544.0     4.434992               4.427040                4.788000     4.582336               4.571872                4.945728       4.015280        3.982320
-5   4608.0     4.499456               4.490816                4.874464     4.669152               4.663648                5.035232       4.071872        3.983520
+0   4224.0     3.801472               3.790064                4.182320     3.968000               3.957824                4.311152       3.398160        3.330400
+1   4352.0     4.082944               4.082912                4.413488     4.400000               4.391936                4.738048       3.837424        3.758208
+2   4416.0     4.142624               4.135648                4.484160     4.452304               4.446096                4.792480       3.892064        3.864128
+3   4480.0     4.206144               4.198752                4.551808     4.530752               4.522944                4.873760       3.949344        3.870224
+4   4544.0     4.438320               4.433104                4.787584     4.584160               4.576640                4.934304       4.008960        3.974672
+5   4608.0     4.502432               4.495456                4.871872     4.660192               4.651040                5.029792       4.065616        3.984160
 </div>
 <div class="uv-install-logs" id="uv-logs-benchmark">
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
-Downloading hf-xet (3.0MiB)
+Downloading networkx (1.9MiB)
+Downloading numpy (16.2MiB)
+Downloading pillow (6.3MiB)
+Downloading fonttools (4.7MiB)
 Downloading setuptools (1.1MiB)
 Downloading nvidia-cufile-cu12 (1.1MiB)
-Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading sympy (6.0MiB)
 Downloading pandas (11.8MiB)
+Downloading hf-xet (3.0MiB)
 Downloading nvidia-cusparse-cu12 (274.9MiB)
-Downloading networkx (1.9MiB)
 Downloading nvidia-cusolver-cu12 (255.1MiB)
-Downloading nvidia-cusparselt-cu12 (273.9MiB)
-Downloading nvidia-cufft-cu12 (184.2MiB)
-Downloading triton (148.3MiB)
+Downloading nvidia-nccl-cu12 (307.4MiB)
 Downloading matplotlib (8.3MiB)
+Downloading nvidia-cufft-cu12 (184.2MiB)
+Downloading nvidia-cusparselt-cu12 (273.9MiB)
+Downloading nvidia-cublas-cu12 (566.8MiB)
+Downloading nvidia-nvjitlink-cu12 (37.4MiB)
 Downloading kiwisolver (1.4MiB)
-Downloading nvidia-nccl-cu12 (307.4MiB)
+Downloading nvidia-cudnn-cu12 (674.0MiB)
+Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
+Downloading torch (846.9MiB)
 Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
 Downloading nvidia-curand-cu12 (60.7MiB)
-Downloading nvidia-cublas-cu12 (566.8MiB)
-Downloading torch (846.9MiB)
-Downloading numpy (16.2MiB)
-Downloading pillow (6.3MiB)
-Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
-Downloading nvidia-cudnn-cu12 (674.0MiB)
-Downloading fonttools (4.7MiB)
-Downloading sympy (6.0MiB)
+Downloading triton (148.3MiB)
  Downloading nvidia-cufile-cu12
  Downloading kiwisolver
  Downloading hf-xet
@@ -4616,34 +4612,29 @@ Downloading sympy (6.0MiB)
  Downloading numpy
  Downloading nvidia-nvjitlink-cu12
  Downloading nvidia-curand-cu12
- Downloading pandas
  Downloading nvidia-cuda-nvrtc-cu12
+ Downloading pandas
  Downloading triton
  Downloading nvidia-cufft-cu12
  Downloading nvidia-cusolver-cu12
- Downloading nvidia-cusparselt-cu12
  Downloading nvidia-cusparse-cu12
+ Downloading nvidia-cusparselt-cu12
  Downloading nvidia-nccl-cu12
  Downloading nvidia-cublas-cu12
  Downloading nvidia-cudnn-cu12
  Downloading torch
-Installed 49 packages in 520ms
+Installed 49 packages in 518ms
 </div>
 </div>
 <div class="cell-stderr">Fetching 20 files:   0%|          | 0/20 [00:00&lt;?, ?it/s]
-Fetching 20 files:   5%|▌         | 1/20 [00:00&lt;00:03,  4.89it/s]
-Fetching 20 files:  10%|█         | 2/20 [00:01&lt;00:17,  1.02it/s]
-Fetching 20 files: 100%|██████████| 20/20 [00:01&lt;00:00, 11.55it/s]
+Fetching 20 files:   5%|▌         | 1/20 [00:00&lt;00:03,  5.10it/s]
+Fetching 20 files:  10%|█         | 2/20 [00:01&lt;00:14,  1.23it/s]
+Fetching 20 files: 100%|██████████| 20/20 [00:01&lt;00:00, 13.86it/s]
 
 Fetching 4 files:   0%|          | 0/4 [00:00&lt;?, ?it/s]
-Fetching 4 files:  25%|██▌       | 1/4 [00:00&lt;00:00,  6.41it/s]
-Fetching 4 files:  50%|█████     | 2/4 [00:01&lt;00:01,  1.10it/s]
-Fetching 4 files: 100%|██████████| 4/4 [00:01&lt;00:00,  2.52it/s]
-Traceback (most recent call last):
-  File &quot;/repo/flash_attn/.uvnote/cells/benchmark.py&quot;, line 344, in &lt;module&gt;
-    fig.savefig(output_dir / &quot;attention_benchmark.png&quot;, dpi=300, bbox_inches=&quot;tight&quot;)
-    ^^^^^^^^^^^
-AttributeError: &#x27;NoneType&#x27; object has no attribute &#x27;savefig&#x27;</div>
+Fetching 4 files:  25%|██▌       | 1/4 [00:00&lt;00:00,  6.31it/s]
+Fetching 4 files:  50%|███��█     | 2/4 [00:01&lt;00:01,  1.34it/s]
+Fetching 4 files: 100%|██████████| 4/4 [00:01&lt;00:00,  3.05it/s]</div>
 <div class="cell-artifacts">
 <h4>Artifacts:</h4>
 <a href="artifacts/benchmark/Attention Benchmark.png" class="artifact" target="_blank">Attention Benchmark.png</a>

flash_attn/cells/benchmark.py

@@ -341,5 +341,3 @@ def benchmark_fn(seq_len: int, provider: str):
 with torch.inference_mode():
     correctness()
     fig = benchmark_fn.run(print_data=True, save_path=output_dir.as_posix())
-    fig.savefig(output_dir / "attention_benchmark.png", dpi=300, bbox_inches="tight")
-    print(f"Benchmark plot saved to: {output_dir / 'attention_benchmark.png'}")

moe_benchmarks/megablocks/megablocks_only.html

@@ -3727,7 +3727,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <span onclick="toggleOutput('forward_and_backward_no_kernel')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-forward_and_backward_no_kernel" style="cursor: default; opacity: 0.3;">▶ uv-logs</span>
 </span> | 
-Cell: forward_and_backward_no_kernel | 17.31s | FAILED
+Cell: forward_and_backward_no_kernel | 17.10s | FAILED
  | <button class="run-btn" onclick="runCell('forward_and_backward_no_kernel')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('forward_and_backward_no_kernel')">Copy</button>
 <a href="cells/forward_and_backward_no_kernel.py" target="_blank" class="raw-btn">Raw</a>

moe_benchmarks/megablocks_yamoe/artifacts/binned_run/binned_results.json

@@ -9,16 +9,16 @@
     "vary_inputs": true
   },
   "stats": {
-    "avg_ms": 37.042515599997614,
-    "min_ms": 33.61098199997059,
-    "max_ms": 39.77627800003347,
-    "std_ms": 1.6558189449135647,
-    "p50_ms": 37.082583499994826,
-    "p95_ms": 39.325366850013665,
-    "p99_ms": 39.73607153999694,
+    "avg_ms": 35.79408963999981,
+    "min_ms": 33.22658100000808,
+    "max_ms": 37.58223699998098,
+    "std_ms": 1.260985811405264,
+    "p50_ms": 36.03647150001166,
+    "p95_ms": 37.377484250018256,
+    "p99_ms": 37.52526078000528,
     "num_iters": 50,
-    "tokens_per_s": 2699.600671829276,
-    "throughput_variance": 122.63980223025922
+    "tokens_per_s": 2793.7573215509365,
+    "throughput_variance": 99.68321642463675
   },
   "output_sum": 3.97190523147583
 }

moe_benchmarks/megablocks_yamoe/artifacts/gptoss_run/gptoss_results.json

@@ -9,16 +9,16 @@
     "vary_inputs": true
   },
   "stats": {
-    "avg_ms": 47.38012357999878,
-    "min_ms": 40.92212500000869,
-    "max_ms": 51.281423999967046,
-    "std_ms": 2.9172375790717613,
-    "p50_ms": 48.13728099998116,
-    "p95_ms": 51.063823949996845,
-    "p99_ms": 51.260956209974324,
+    "avg_ms": 45.9334355199951,
+    "min_ms": 40.05551199998081,
+    "max_ms": 49.51232600001276,
+    "std_ms": 2.4709340263031536,
+    "p50_ms": 46.49940249998963,
+    "p95_ms": 49.05830289997937,
+    "p99_ms": 49.3528599099983,
     "num_iters": 50,
-    "tokens_per_s": 2110.589682847817,
-    "throughput_variance": 134.12269492084684
+    "tokens_per_s": 2177.0633715492368,
+    "throughput_variance": 121.25434497483073
   },
   "output_sum": 11.53223705291748
 }

moe_benchmarks/megablocks_yamoe/artifacts/gptoss_training_run/gptoss_training_results.json

@@ -9,16 +9,16 @@
     "vary_inputs": true
   },
   "stats": {
-    "avg_ms": 46.92225349999944,
-    "min_ms": 38.89427600000772,
-    "max_ms": 51.62209400003803,
-    "std_ms": 3.930283839179673,
-    "p50_ms": 48.18643950000023,
-    "p95_ms": 51.4210894500053,
-    "p99_ms": 51.56389033003563,
+    "avg_ms": 45.94743435999817,
+    "min_ms": 38.690121999991334,
+    "max_ms": 51.193351999984316,
+    "std_ms": 3.91507100876056,
+    "p50_ms": 45.20909099997539,
+    "p95_ms": 51.039028550002286,
+    "p99_ms": 51.14429515998495,
     "num_iters": 50,
-    "tokens_per_s": 2131.1849397855794,
-    "throughput_variance": 188.79708542409617
+    "tokens_per_s": 2176.4000839851024,
+    "throughput_variance": 188.75969966024954
   },
   "output_sum": 11.53223705291748
 }

moe_benchmarks/megablocks_yamoe/artifacts/megablocks_run/megablocks_results.json

@@ -9,16 +9,16 @@
     "vary_inputs": true
   },
   "stats": {
-    "avg_ms": 3.8653191399976095,
-    "min_ms": 0.8422269999073251,
-    "max_ms": 8.544625000013184,
-    "std_ms": 3.690530253469649,
-    "p50_ms": 0.88288749992671,
-    "p95_ms": 8.536876499982782,
-    "p99_ms": 8.542453809967583,
+    "avg_ms": 3.8478457200017147,
+    "min_ms": 0.8121239999354657,
+    "max_ms": 8.535666000057063,
+    "std_ms": 3.697659288553723,
+    "p50_ms": 0.8394504999955643,
+    "p95_ms": 8.499624499950187,
+    "p99_ms": 8.528520820026415,
     "num_iters": 50,
-    "tokens_per_s": 25871.084994048342,
-    "throughput_variance": 50674.824252369
+    "tokens_per_s": 25988.567961595778,
+    "throughput_variance": 53035.39729321811
   },
   "output_sum": 6.4738850593566895
 }

moe_benchmarks/megablocks_yamoe/artifacts/yamoe_run/yamoe_results.json

@@ -9,16 +9,16 @@
     "vary_inputs": true
   },
   "stats": {
-    "avg_ms": 4.248850420003691,
-    "min_ms": 4.146223000020655,
-    "max_ms": 4.269965999981196,
-    "std_ms": 0.01914249322297606,
-    "p50_ms": 4.252545499980442,
-    "p95_ms": 4.265578499993694,
-    "p99_ms": 4.269833699987657,
+    "avg_ms": 4.246394759998111,
+    "min_ms": 4.066528999999264,
+    "max_ms": 4.294285000014497,
+    "std_ms": 0.033808054217192726,
+    "p50_ms": 4.2530110000313925,
+    "p95_ms": 4.267295049984909,
+    "p99_ms": 4.287134920007816,
     "num_iters": 50,
-    "tokens_per_s": 23535.77794341678,
-    "throughput_variance": 107.68667127056374
+    "tokens_per_s": 23549.38851705923,
+    "throughput_variance": 193.18069406896424
   },
-  "output_sum": 3.9719059467315674
+  "output_sum": 3.97190523147583
 }

moe_benchmarks/megablocks_yamoe/megablocks_yamoe.html

@@ -3726,7 +3726,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <span onclick="toggleOutput('setup')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-setup" style="cursor: default; opacity: 0.3;">▶ uv-logs</span>
 </span> | 
-Cell: setup | 17.08s | FAILED
+Cell: setup | 16.96s | FAILED
  | <button class="run-btn" onclick="runCell('setup')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('setup')">Copy</button>
 <a href="cells/setup.py" target="_blank" class="raw-btn">Raw</a>

moe_benchmarks/megablocks_yamoe/torch_profile.html

@@ -3720,7 +3720,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <span onclick="toggleOutput('utils')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-utils" onclick="toggleUvLogsFromHeader('utils')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: utils | deps: torch, numpy | 34.86s
+Cell: utils | deps: torch, numpy | 34.47s
  | <button class="run-btn" onclick="runCell('utils')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('utils')">Copy</button>
 <a href="cells/utils.py" target="_blank" class="raw-btn">Raw</a>
@@ -3794,24 +3794,24 @@ Cell: utils | deps: torch, numpy | 34.86s
 <div class="uv-install-logs" id="uv-logs-utils">
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
-Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
-Downloading sympy (6.0MiB)
 Downloading networkx (1.9MiB)
-Downloading setuptools (1.1MiB)
-Downloading nvidia-cudnn-cu12 (674.0MiB)
-Downloading nvidia-cufile-cu12 (1.1MiB)
 Downloading nvidia-nccl-cu12 (307.4MiB)
+Downloading setuptools (1.1MiB)
+Downloading numpy (16.2MiB)
+Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
+Downloading sympy (6.0MiB)
 Downloading nvidia-cusolver-cu12 (255.1MiB)
-Downloading nvidia-cusparselt-cu12 (273.9MiB)
 Downloading nvidia-curand-cu12 (60.7MiB)
-Downloading nvidia-nvjitlink-cu12 (37.4MiB)
-Downloading numpy (16.2MiB)
-Downloading triton (148.3MiB)
-Downloading nvidia-cublas-cu12 (566.8MiB)
+Downloading nvidia-cudnn-cu12 (674.0MiB)
+Downloading torch (846.9MiB)
 Downloading nvidia-cufft-cu12 (184.2MiB)
+Downloading nvidia-cublas-cu12 (566.8MiB)
 Downloading nvidia-cusparse-cu12 (274.9MiB)
-Downloading torch (846.9MiB)
-Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
+Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading nvidia-cusparselt-cu12 (273.9MiB)
+Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
+Downloading triton (148.3MiB)
+Downloading nvidia-cufile-cu12 (1.1MiB)
  Downloading nvidia-cufile-cu12
  Downloading setuptools
  Downloading networkx
@@ -3843,7 +3843,7 @@ Installed 26 packages in 447ms
 <span onclick="toggleOutput('bench_utils')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-bench_utils" onclick="toggleUvLogsFromHeader('bench_utils')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: bench_utils | deps: torch, numpy | 35.19s
+Cell: bench_utils | deps: torch, numpy | 34.94s
  | <button class="run-btn" onclick="runCell('bench_utils')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('bench_utils')">Copy</button>
 <a href="cells/bench_utils.py" target="_blank" class="raw-btn">Raw</a>
@@ -4331,24 +4331,24 @@ Cell: bench_utils | deps: torch, numpy | 35.19s
 <div class="uv-install-logs" id="uv-logs-bench_utils">
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
-Downloading sympy (6.0MiB)
-Downloading networkx (1.9MiB)
 Downloading setuptools (1.1MiB)
-Downloading nvidia-cudnn-cu12 (674.0MiB)
-Downloading nvidia-cusolver-cu12 (255.1MiB)
+Downloading sympy (6.0MiB)
+Downloading nvidia-curand-cu12 (60.7MiB)
+Downloading nvidia-cusparse-cu12 (274.9MiB)
 Downloading numpy (16.2MiB)
-Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
-Downloading nvidia-nvjitlink-cu12 (37.4MiB)
-Downloading torch (846.9MiB)
 Downloading nvidia-cufft-cu12 (184.2MiB)
-Downloading nvidia-cusparse-cu12 (274.9MiB)
+Downloading networkx (1.9MiB)
+Downloading nvidia-nvjitlink-cu12 (37.4MiB)
 Downloading nvidia-cublas-cu12 (566.8MiB)
-Downloading nvidia-curand-cu12 (60.7MiB)
 Downloading nvidia-cufile-cu12 (1.1MiB)
+Downloading torch (846.9MiB)
+Downloading nvidia-cudnn-cu12 (674.0MiB)
+Downloading nvidia-nccl-cu12 (307.4MiB)
+Downloading nvidia-cusolver-cu12 (255.1MiB)
+Downloading triton (148.3MiB)
 Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
 Downloading nvidia-cusparselt-cu12 (273.9MiB)
-Downloading triton (148.3MiB)
-Downloading nvidia-nccl-cu12 (307.4MiB)
+Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
  Downloading nvidia-cufile-cu12
  Downloading setuptools
  Downloading networkx
@@ -4361,8 +4361,8 @@ Downloading nvidia-nccl-cu12 (307.4MiB)
  Downloading triton
  Downloading nvidia-cufft-cu12
  Downloading nvidia-cusolver-cu12
- Downloading nvidia-cusparselt-cu12
  Downloading nvidia-cusparse-cu12
+ Downloading nvidia-cusparselt-cu12
  Downloading nvidia-nccl-cu12
  Downloading nvidia-cublas-cu12
  Downloading nvidia-cudnn-cu12
@@ -4381,7 +4381,7 @@ Installed 26 packages in 455ms
 <span onclick="toggleOutput('config')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-config" onclick="toggleUvLogsFromHeader('config')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: config | deps: torch, numpy | 34.96s
+Cell: config | deps: torch, numpy | 35.62s
  | <button class="run-btn" onclick="runCell('config')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('config')">Copy</button>
 <a href="cells/config.py" target="_blank" class="raw-btn">Raw</a>
@@ -4441,24 +4441,24 @@ Cell: config | deps: torch, numpy | 34.96s
 <div class="uv-install-logs" id="uv-logs-config">
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
+Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading sympy (6.0MiB)
 Downloading nvidia-cufile-cu12 (1.1MiB)
+Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
+Downloading setuptools (1.1MiB)
+Downloading networkx (1.9MiB)
+Downloading nvidia-cublas-cu12 (566.8MiB)
 Downloading nvidia-cusparse-cu12 (274.9MiB)
+Downloading nvidia-nccl-cu12 (307.4MiB)
+Downloading nvidia-cusolver-cu12 (255.1MiB)
 Downloading nvidia-cufft-cu12 (184.2MiB)
-Downloading sympy (6.0MiB)
-Downloading triton (148.3MiB)
 Downloading nvidia-curand-cu12 (60.7MiB)
+Downloading numpy (16.2MiB)
+Downloading nvidia-cudnn-cu12 (674.0MiB)
 Downloading nvidia-cusparselt-cu12 (273.9MiB)
-Downloading nvidia-cublas-cu12 (566.8MiB)
-Downloading nvidia-nvjitlink-cu12 (37.4MiB)
-Downloading setuptools (1.1MiB)
-Downloading nvidia-nccl-cu12 (307.4MiB)
-Downloading networkx (1.9MiB)
 Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
-Downloading nvidia-cusolver-cu12 (255.1MiB)
-Downloading nvidia-cudnn-cu12 (674.0MiB)
-Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
 Downloading torch (846.9MiB)
-Downloading numpy (16.2MiB)
+Downloading triton (148.3MiB)
  Downloading nvidia-cufile-cu12
  Downloading setuptools
  Downloading networkx
@@ -4471,8 +4471,8 @@ Downloading numpy (16.2MiB)
  Downloading triton
  Downloading nvidia-cufft-cu12
  Downloading nvidia-cusolver-cu12
- Downloading nvidia-cusparse-cu12
  Downloading nvidia-cusparselt-cu12
+ Downloading nvidia-cusparse-cu12
  Downloading nvidia-nccl-cu12
  Downloading nvidia-cublas-cu12
  Downloading nvidia-cudnn-cu12
@@ -4490,7 +4490,7 @@ Installed 26 packages in 449ms
 <span onclick="toggleOutput('save_data')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-save_data" onclick="toggleUvLogsFromHeader('save_data')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: save_data | deps: torch, numpy | 39.26s
+Cell: save_data | deps: torch, numpy | 39.76s
  | <button class="run-btn" onclick="runCell('save_data')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('save_data')">Copy</button>
 <a href="cells/save_data.py" target="_blank" class="raw-btn">Raw</a>
@@ -4585,24 +4585,24 @@ Down sum: 206.729263
 <div class="uv-install-logs" id="uv-logs-save_data">
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
-Downloading setuptools (1.1MiB)
-Downloading nvidia-cudnn-cu12 (674.0MiB)
-Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
+Downloading nvidia-cufile-cu12 (1.1MiB)
 Downloading nvidia-nccl-cu12 (307.4MiB)
-Downloading nvidia-cusparse-cu12 (274.9MiB)
 Downloading sympy (6.0MiB)
+Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
+Downloading nvidia-cufft-cu12 (184.2MiB)
+Downloading numpy (16.2MiB)
 Downloading nvidia-cusolver-cu12 (255.1MiB)
+Downloading nvidia-cusparse-cu12 (274.9MiB)
 Downloading nvidia-curand-cu12 (60.7MiB)
-Downloading numpy (16.2MiB)
-Downloading nvidia-cufile-cu12 (1.1MiB)
-Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
 Downloading nvidia-cusparselt-cu12 (273.9MiB)
-Downloading triton (148.3MiB)
-Downloading nvidia-cufft-cu12 (184.2MiB)
+Downloading setuptools (1.1MiB)
+Downloading nvidia-cudnn-cu12 (674.0MiB)
+Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
 Downloading networkx (1.9MiB)
-Downloading torch (846.9MiB)
 Downloading nvidia-cublas-cu12 (566.8MiB)
+Downloading triton (148.3MiB)
 Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading torch (846.9MiB)
  Downloading nvidia-cufile-cu12
  Downloading setuptools
  Downloading networkx
@@ -4621,17 +4621,17 @@ Downloading nvidia-nvjitlink-cu12 (37.4MiB)
  Downloading nvidia-cublas-cu12
  Downloading nvidia-cudnn-cu12
  Downloading torch
-Installed 26 packages in 453ms
+Installed 26 packages in 447ms
 </div>
 </div>
 <div class="cell-artifacts">
 <h4>Artifacts:</h4>
 <a href="artifacts/save_data/router_weight.pt" class="artifact" target="_blank">router_weight.pt</a>
 <a href="artifacts/save_data/gate_up_proj_bias.pt" class="artifact" target="_blank">gate_up_proj_bias.pt</a>
-<a href="artifacts/save_data/gate_up_proj.pt" class="artifact" target="_blank">gate_up_proj.pt</a>
-<a href="artifacts/save_data/down_proj.pt" class="artifact" target="_blank">down_proj.pt</a>
 <a href="artifacts/save_data/router_bias.pt" class="artifact" target="_blank">router_bias.pt</a>
+<a href="artifacts/save_data/gate_up_proj.pt" class="artifact" target="_blank">gate_up_proj.pt</a>
 <a href="artifacts/save_data/down_proj_bias.pt" class="artifact" target="_blank">down_proj_bias.pt</a>
+<a href="artifacts/save_data/down_proj.pt" class="artifact" target="_blank">down_proj.pt</a>
 </div>
 </div>
 </div>
@@ -4645,7 +4645,7 @@ Installed 26 packages in 453ms
 <span onclick="toggleOutput('yamoe_run')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-yamoe_run" onclick="toggleUvLogsFromHeader('yamoe_run')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: yamoe_run | deps: torch, kernels, numpy | 39.67s
+Cell: yamoe_run | deps: torch, kernels, numpy | 38.79s
  | <button class="run-btn" onclick="runCell('yamoe_run')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('yamoe_run')">Copy</button>
 <a href="cells/yamoe_run.py" target="_blank" class="raw-btn">Raw</a>
@@ -4938,10 +4938,10 @@ Input Variation: +0.001 * iteration (deterministic)
 
 Warming up (10 iterations)...
 Benchmarking (50 iterations)...
-  Progress: 20% complete (avg: 4.250 ms)
-  Progress: 40% complete (avg: 4.247 ms)
-  Progress: 60% complete (avg: 4.248 ms)
-  Progress: 80% complete (avg: 4.248 ms)
+  Progress: 20% complete (avg: 4.251 ms)
+  Progress: 40% complete (avg: 4.249 ms)
+  Progress: 60% complete (avg: 4.244 ms)
+  Progress: 80% complete (avg: 4.246 ms)
 
 Output tensors:
   Primary: shape=(1, 100, 1152), dtype=torch.float32, device=cuda:0, range=[-0.049506, 0.054984], mean=0.000034, std=0.006508, norm=2.208791
@@ -4951,47 +4951,47 @@ Output tensors:
 Iterations: 50
 
 Latency Statistics:
-  Average: 4.249 ms
-  Min:     4.146 ms
-  Max:     4.270 ms
-  Std Dev: 0.019 ms
+  Average: 4.246 ms
+  Min:     4.067 ms
+  Max:     4.294 ms
+  Std Dev: 0.034 ms
 
 Percentiles:
   P50 (median): 4.253 ms
-  P95:          4.266 ms
-  P99:          4.270 ms
+  P95:          4.267 ms
+  P99:          4.287 ms
 
 Throughput:
-  Tokens/sec: 23535.8
-  Std Dev:    107.7
+  Tokens/sec: 23549.4
+  Std Dev:    193.2
 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 
 Saved benchmark results to yamoe_results.json
 
-Output sum: 3.971906
+Output sum: 3.971905
 </div>
 <div class="uv-install-logs" id="uv-logs-yamoe_run">
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
-Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
-Downloading networkx (1.9MiB)
-Downloading nvidia-cufft-cu12 (184.2MiB)
-Downloading nvidia-cudnn-cu12 (674.0MiB)
+Downloading numpy (16.2MiB)
+Downloading nvidia-cufile-cu12 (1.1MiB)
 Downloading nvidia-cusparselt-cu12 (273.9MiB)
-Downloading nvidia-cublas-cu12 (566.8MiB)
 Downloading torch (846.9MiB)
-Downloading nvidia-cusparse-cu12 (274.9MiB)
-Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
-Downloading nvidia-nccl-cu12 (307.4MiB)
+Downloading nvidia-cufft-cu12 (184.2MiB)
 Downloading nvidia-cusolver-cu12 (255.1MiB)
-Downloading triton (148.3MiB)
-Downloading numpy (16.2MiB)
+Downloading nvidia-nccl-cu12 (307.4MiB)
+Downloading setuptools (1.1MiB)
 Downloading nvidia-curand-cu12 (60.7MiB)
+Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
+Downloading nvidia-cublas-cu12 (566.8MiB)
 Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading nvidia-cusparse-cu12 (274.9MiB)
+Downloading networkx (1.9MiB)
 Downloading sympy (6.0MiB)
-Downloading setuptools (1.1MiB)
+Downloading nvidia-cudnn-cu12 (674.0MiB)
 Downloading hf-xet (3.0MiB)
-Downloading nvidia-cufile-cu12 (1.1MiB)
+Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
+Downloading triton (148.3MiB)
  Downloading nvidia-cufile-cu12
  Downloading hf-xet
  Downloading setuptools
@@ -5005,19 +5005,19 @@ Downloading nvidia-cufile-cu12 (1.1MiB)
  Downloading triton
  Downloading nvidia-cufft-cu12
  Downloading nvidia-cusolver-cu12
- Downloading nvidia-cusparselt-cu12
  Downloading nvidia-cusparse-cu12
+ Downloading nvidia-cusparselt-cu12
  Downloading nvidia-nccl-cu12
  Downloading nvidia-cublas-cu12
  Downloading nvidia-cudnn-cu12
  Downloading torch
-Installed 37 packages in 453ms
+Installed 37 packages in 449ms
 </div>
 </div>
 <div class="cell-stderr">Fetching 6 files:   0%|          | 0/6 [00:00&lt;?, ?it/s]
-Fetching 6 files:  17%|█▋        | 1/6 [00:00&lt;00:01,  3.69it/s]
-Fetching 6 files:  50%|█████     | 3/6 [00:00&lt;00:00,  3.01it/s]
-Fetching 6 files: 100%|██████████| 6/6 [00:00&lt;00:00,  6.13it/s]</div>
+Fetching 6 files:  17%|█▋        | 1/6 [00:00&lt;00:01,  3.75it/s]
+Fetching 6 files:  50%|█████     | 3/6 [00:00&lt;00:00,  3.39it/s]
+Fetching 6 files: 100%|██████████| 6/6 [00:00&lt;00:00,  6.84it/s]</div>
 <div class="cell-artifacts">
 <h4>Artifacts:</h4>
 <a href="artifacts/yamoe_run/yamoe_results.json" class="artifact" target="_blank">yamoe_results.json</a>
@@ -5034,7 +5034,7 @@ Fetching 6 files: 100%|██████████| 6/6 [00:00&lt;00:00,  6.1
 <span onclick="toggleOutput('binned_run')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-binned_run" onclick="toggleUvLogsFromHeader('binned_run')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: binned_run | deps: torch, numpy | 39.33s
+Cell: binned_run | deps: torch, numpy | 39.53s
  | <button class="run-btn" onclick="runCell('binned_run')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('binned_run')">Copy</button>
 <a href="cells/binned_run.py" target="_blank" class="raw-btn">Raw</a>
@@ -5448,10 +5448,10 @@ Input Variation: +0.001 * iteration (deterministic)
 
 Warming up (10 iterations)...
 Benchmarking (50 iterations)...
-  Progress: 20% complete (avg: 38.670 ms)
-  Progress: 40% complete (avg: 38.443 ms)
-  Progress: 60% complete (avg: 38.101 ms)
-  Progress: 80% complete (avg: 37.640 ms)
+  Progress: 20% complete (avg: 37.247 ms)
+  Progress: 40% complete (avg: 37.082 ms)
+  Progress: 60% complete (avg: 36.706 ms)
+  Progress: 80% complete (avg: 36.240 ms)
 
 Output tensors:
   Primary: shape=(1, 100, 1152), dtype=torch.float32, device=cuda:0, range=[-0.049506, 0.054984], mean=0.000034, std=0.006508, norm=2.208791
@@ -5461,19 +5461,19 @@ Output tensors:
 Iterations: 50
 
 Latency Statistics:
-  Average: 37.043 ms
-  Min:     33.611 ms
-  Max:     39.776 ms
-  Std Dev: 1.656 ms
+  Average: 35.794 ms
+  Min:     33.227 ms
+  Max:     37.582 ms
+  Std Dev: 1.261 ms
 
 Percentiles:
-  P50 (median): 37.083 ms
-  P95:          39.325 ms
-  P99:          39.736 ms
+  P50 (median): 36.036 ms
+  P95:          37.377 ms
+  P99:          37.525 ms
 
 Throughput:
-  Tokens/sec: 2699.6
-  Std Dev:    122.6
+  Tokens/sec: 2793.8
+  Std Dev:    99.7
 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 
 Saved benchmark results to binned_results.json
@@ -5483,23 +5483,23 @@ Output sum: 3.971905
 <div class="uv-install-logs" id="uv-logs-binned_run">
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
+Downloading sympy (6.0MiB)
+Downloading nvidia-cufile-cu12 (1.1MiB)
+Downloading nvidia-cudnn-cu12 (674.0MiB)
 Downloading setuptools (1.1MiB)
+Downloading numpy (16.2MiB)
 Downloading nvidia-nvjitlink-cu12 (37.4MiB)
-Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
-Downloading networkx (1.9MiB)
-Downloading nvidia-cublas-cu12 (566.8MiB)
-Downloading nvidia-cufft-cu12 (184.2MiB)
 Downloading triton (148.3MiB)
+Downloading nvidia-cufft-cu12 (184.2MiB)
 Downloading nvidia-nccl-cu12 (307.4MiB)
-Downloading nvidia-cusolver-cu12 (255.1MiB)
+Downloading nvidia-cublas-cu12 (566.8MiB)
 Downloading torch (846.9MiB)
+Downloading nvidia-cusolver-cu12 (255.1MiB)
+Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
 Downloading nvidia-curand-cu12 (60.7MiB)
-Downloading numpy (16.2MiB)
+Downloading networkx (1.9MiB)
 Downloading nvidia-cusparselt-cu12 (273.9MiB)
-Downloading sympy (6.0MiB)
 Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
-Downloading nvidia-cufile-cu12 (1.1MiB)
-Downloading nvidia-cudnn-cu12 (674.0MiB)
 Downloading nvidia-cusparse-cu12 (274.9MiB)
  Downloading nvidia-cufile-cu12
  Downloading setuptools
@@ -5513,13 +5513,13 @@ Downloading nvidia-cusparse-cu12 (274.9MiB)
  Downloading triton
  Downloading nvidia-cufft-cu12
  Downloading nvidia-cusolver-cu12
- Downloading nvidia-cusparse-cu12
  Downloading nvidia-cusparselt-cu12
+ Downloading nvidia-cusparse-cu12
  Downloading nvidia-nccl-cu12
  Downloading nvidia-cublas-cu12
  Downloading nvidia-cudnn-cu12
  Downloading torch
-Installed 26 packages in 526ms
+Installed 26 packages in 525ms
 </div>
 </div>
 <div class="cell-artifacts">
@@ -5538,7 +5538,7 @@ Installed 26 packages in 526ms
 <span onclick="toggleOutput('gptoss_run')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-gptoss_run" onclick="toggleUvLogsFromHeader('gptoss_run')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: gptoss_run | deps: torch, numpy | 40.94s
+Cell: gptoss_run | deps: torch, numpy | 40.29s
  | <button class="run-btn" onclick="runCell('gptoss_run')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('gptoss_run')">Copy</button>
 <a href="cells/gptoss_run.py" target="_blank" class="raw-btn">Raw</a>
@@ -5856,10 +5856,10 @@ Input Variation: +0.001 * iteration (deterministic)
 
 Warming up (10 iterations)...
 Benchmarking (50 iterations)...
-  Progress: 20% complete (avg: 50.824 ms)
-  Progress: 40% complete (avg: 50.316 ms)
-  Progress: 60% complete (avg: 49.490 ms)
-  Progress: 80% complete (avg: 48.422 ms)
+  Progress: 20% complete (avg: 48.814 ms)
+  Progress: 40% complete (avg: 48.182 ms)
+  Progress: 60% complete (avg: 47.686 ms)
+  Progress: 80% complete (avg: 46.880 ms)
 
 Output tensors:
   Primary: shape=(1, 100, 1152), dtype=torch.float32, device=cuda:0, range=[-0.064982, 0.061193], mean=0.000100, std=0.013510, norm=4.585560
@@ -5869,19 +5869,19 @@ Output tensors:
 Iterations: 50
 
 Latency Statistics:
-  Average: 47.380 ms
-  Min:     40.922 ms
-  Max:     51.281 ms
-  Std Dev: 2.917 ms
+  Average: 45.933 ms
+  Min:     40.056 ms
+  Max:     49.512 ms
+  Std Dev: 2.471 ms
 
 Percentiles:
-  P50 (median): 48.137 ms
-  P95:          51.064 ms
-  P99:          51.261 ms
+  P50 (median): 46.499 ms
+  P95:          49.058 ms
+  P99:          49.353 ms
 
 Throughput:
-  Tokens/sec: 2110.6
-  Std Dev:    134.1
+  Tokens/sec: 2177.1
+  Std Dev:    121.3
 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━��━━━━━━━━━━━━━
 
 Saved benchmark results to gptoss_results.json
@@ -5892,23 +5892,23 @@ Output sum: 11.532237
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
 Downloading setuptools (1.1MiB)
+Downloading sympy (6.0MiB)
+Downloading numpy (16.2MiB)
 Downloading networkx (1.9MiB)
-Downloading triton (148.3MiB)
-Downloading nvidia-cufile-cu12 (1.1MiB)
-Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
-Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading nvidia-cusolver-cu12 (255.1MiB)
+Downloading nvidia-cusparselt-cu12 (273.9MiB)
+Downloading nvidia-cusparse-cu12 (274.9MiB)
+Downloading nvidia-cublas-cu12 (566.8MiB)
+Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
 Downloading nvidia-cufft-cu12 (184.2MiB)
-Downloading nvidia-cudnn-cu12 (674.0MiB)
 Downloading nvidia-curand-cu12 (60.7MiB)
-Downloading numpy (16.2MiB)
 Downloading nvidia-nccl-cu12 (307.4MiB)
-Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
-Downloading nvidia-cusolver-cu12 (255.1MiB)
-Downloading sympy (6.0MiB)
-Downloading nvidia-cusparse-cu12 (274.9MiB)
+Downloading nvidia-cufile-cu12 (1.1MiB)
+Downloading nvidia-cudnn-cu12 (674.0MiB)
 Downloading torch (846.9MiB)
-Downloading nvidia-cusparselt-cu12 (273.9MiB)
-Downloading nvidia-cublas-cu12 (566.8MiB)
+Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
+Downloading triton (148.3MiB)
  Downloading nvidia-cufile-cu12
  Downloading setuptools
  Downloading networkx
@@ -5927,7 +5927,7 @@ Downloading nvidia-cublas-cu12 (566.8MiB)
  Downloading nvidia-cublas-cu12
  Downloading nvidia-cudnn-cu12
  Downloading torch
-Installed 26 packages in 449ms
+Installed 26 packages in 540ms
 </div>
 </div>
 <div class="cell-artifacts">
@@ -5946,7 +5946,7 @@ Installed 26 packages in 449ms
 <span onclick="toggleOutput('gptoss_training_run')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-gptoss_training_run" onclick="toggleUvLogsFromHeader('gptoss_training_run')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: gptoss_training_run | deps: torch, numpy | 40.19s
+Cell: gptoss_training_run | deps: torch, numpy | 39.76s
  | <button class="run-btn" onclick="runCell('gptoss_training_run')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('gptoss_training_run')">Copy</button>
 <a href="cells/gptoss_training_run.py" target="_blank" class="raw-btn">Raw</a>
@@ -6247,10 +6247,10 @@ Input Variation: +0.001 * iteration (deterministic)
 
 Warming up (10 iterations)...
 Benchmarking (50 iterations)...
-  Progress: 20% complete (avg: 51.217 ms)
-  Progress: 40% complete (avg: 50.712 ms)
-  Progress: 60% complete (avg: 49.798 ms)
-  Progress: 80% complete (avg: 48.389 ms)
+  Progress: 20% complete (avg: 50.744 ms)
+  Progress: 40% complete (avg: 50.240 ms)
+  Progress: 60% complete (avg: 48.683 ms)
+  Progress: 80% complete (avg: 47.222 ms)
 
 Output tensors:
   Primary: shape=(1, 100, 1152), dtype=torch.float32, device=cuda:0, range=[-0.064982, 0.061193], mean=0.000100, std=0.013510, norm=4.585560
@@ -6260,18 +6260,18 @@ Output tensors:
 Iterations: 50
 
 Latency Statistics:
-  Average: 46.922 ms
-  Min:     38.894 ms
-  Max:     51.622 ms
-  Std Dev: 3.930 ms
+  Average: 45.947 ms
+  Min:     38.690 ms
+  Max:     51.193 ms
+  Std Dev: 3.915 ms
 
 Percentiles:
-  P50 (median): 48.186 ms
-  P95:          51.421 ms
-  P99:          51.564 ms
+  P50 (median): 45.209 ms
+  P95:          51.039 ms
+  P99:          51.144 ms
 
 Throughput:
-  Tokens/sec: 2131.2
+  Tokens/sec: 2176.4
   Std Dev:    188.8
 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 
@@ -6282,23 +6282,23 @@ Output sum: 11.532237
 <div class="uv-install-logs" id="uv-logs-gptoss_training_run">
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
-Downloading nvidia-cudnn-cu12 (674.0MiB)
-Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
-Downloading nvidia-cusolver-cu12 (255.1MiB)
-Downloading nvidia-nccl-cu12 (307.4MiB)
 Downloading nvidia-cufft-cu12 (184.2MiB)
-Downloading nvidia-cusparse-cu12 (274.9MiB)
-Downloading networkx (1.9MiB)
-Downloading nvidia-curand-cu12 (60.7MiB)
+Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
 Downloading sympy (6.0MiB)
-Downloading numpy (16.2MiB)
 Downloading nvidia-cufile-cu12 (1.1MiB)
-Downloading nvidia-cusparselt-cu12 (273.9MiB)
 Downloading torch (846.9MiB)
+Downloading nvidia-cusparse-cu12 (274.9MiB)
 Downloading setuptools (1.1MiB)
-Downloading nvidia-nvjitlink-cu12 (37.4MiB)
-Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
+Downloading nvidia-curand-cu12 (60.7MiB)
+Downloading networkx (1.9MiB)
+Downloading numpy (16.2MiB)
+Downloading nvidia-cudnn-cu12 (674.0MiB)
 Downloading nvidia-cublas-cu12 (566.8MiB)
+Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
+Downloading nvidia-cusparselt-cu12 (273.9MiB)
+Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading nvidia-nccl-cu12 (307.4MiB)
+Downloading nvidia-cusolver-cu12 (255.1MiB)
 Downloading triton (148.3MiB)
  Downloading nvidia-cufile-cu12
  Downloading setuptools
@@ -6318,7 +6318,7 @@ Downloading triton (148.3MiB)
  Downloading nvidia-cublas-cu12
  Downloading nvidia-cudnn-cu12
  Downloading torch
-Installed 26 packages in 548ms
+Installed 26 packages in 444ms
 </div>
 </div>
 <div class="cell-artifacts">
@@ -6337,7 +6337,7 @@ Installed 26 packages in 548ms
 <span onclick="toggleOutput('megablocks_run')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-megablocks_run" onclick="toggleUvLogsFromHeader('megablocks_run')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: megablocks_run | deps: torch, numpy, kernels | 48.02s
+Cell: megablocks_run | deps: torch, numpy, kernels | 47.11s
  | <button class="run-btn" onclick="runCell('megablocks_run')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('megablocks_run')">Copy</button>
 <a href="cells/megablocks_run.py" target="_blank" class="raw-btn">Raw</a>
@@ -6566,10 +6566,10 @@ Input Variation: +0.001 * iteration (deterministic)
 
 Warming up (10 iterations)...
 Benchmarking (50 iterations)...
-  Progress: 20% complete (avg: 0.894 ms)
-  Progress: 40% complete (avg: 0.876 ms)
-  Progress: 60% complete (avg: 0.872 ms)
-  Progress: 80% complete (avg: 2.709 ms)
+  Progress: 20% complete (avg: 0.855 ms)
+  Progress: 40% complete (avg: 0.840 ms)
+  Progress: 60% complete (avg: 0.838 ms)
+  Progress: 80% complete (avg: 2.699 ms)
 
 Output tensors:
   Primary: shape=(1, 100, 1152), dtype=torch.float32, device=cuda:0, range=[-0.061104, 0.055115], mean=0.000056, std=0.013535, norm=4.593927
@@ -6579,19 +6579,19 @@ Output tensors:
 Iterations: 50
 
 Latency Statistics:
-  Average: 3.865 ms
-  Min:     0.842 ms
-  Max:     8.545 ms
-  Std Dev: 3.691 ms
+  Average: 3.848 ms
+  Min:     0.812 ms
+  Max:     8.536 ms
+  Std Dev: 3.698 ms
 
 Percentiles:
-  P50 (median): 0.883 ms
-  P95:          8.537 ms
-  P99:          8.542 ms
+  P50 (median): 0.839 ms
+  P95:          8.500 ms
+  P99:          8.529 ms
 
 Throughput:
-  Tokens/sec: 25871.1
-  Std Dev:    50674.8
+  Tokens/sec: 25988.6
+  Std Dev:    53035.4
 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 
 Saved benchmark results to megablocks_results.json
@@ -6602,24 +6602,24 @@ Output sum: 6.473885
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
 Downloading sympy (6.0MiB)
+Downloading nvidia-cudnn-cu12 (674.0MiB)
 Downloading numpy (16.2MiB)
-Downloading nvidia-nccl-cu12 (307.4MiB)
 Downloading hf-xet (3.0MiB)
-Downloading setuptools (1.1MiB)
-Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
-Downloading nvidia-cublas-cu12 (566.8MiB)
-Downloading torch (846.9MiB)
-Downloading nvidia-nvjitlink-cu12 (37.4MiB)
-Downloading nvidia-cufile-cu12 (1.1MiB)
-Downloading nvidia-cudnn-cu12 (674.0MiB)
 Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
+Downloading nvidia-curand-cu12 (60.7MiB)
 Downloading nvidia-cusparse-cu12 (274.9MiB)
-Downloading nvidia-cusparselt-cu12 (273.9MiB)
-Downloading nvidia-cufft-cu12 (184.2MiB)
-Downloading triton (148.3MiB)
+Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading nvidia-cufile-cu12 (1.1MiB)
 Downloading networkx (1.9MiB)
-Downloading nvidia-curand-cu12 (60.7MiB)
+Downloading triton (148.3MiB)
+Downloading nvidia-cufft-cu12 (184.2MiB)
+Downloading nvidia-cusparselt-cu12 (273.9MiB)
 Downloading nvidia-cusolver-cu12 (255.1MiB)
+Downloading torch (846.9MiB)
+Downloading setuptools (1.1MiB)
+Downloading nvidia-cublas-cu12 (566.8MiB)
+Downloading nvidia-nccl-cu12 (307.4MiB)
+Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
  Downloading nvidia-cufile-cu12
  Downloading hf-xet
  Downloading setuptools
@@ -6639,19 +6639,20 @@ Downloading nvidia-cusolver-cu12 (255.1MiB)
  Downloading nvidia-cublas-cu12
  Downloading nvidia-cudnn-cu12
  Downloading torch
-Installed 37 packages in 525ms
+Installed 37 packages in 459ms
 </div>
 </div>
 <div class="cell-stderr">Fetching 66 files:   0%|          | 0/66 [00:00&lt;?, ?it/s]
-Fetching 66 files:   2%|▏         | 1/66 [00:00&lt;00:15,  4.11it/s]
-Fetching 66 files:  15%|█▌        | 10/66 [00:00&lt;00:02, 21.70it/s]
-Fetching 66 files:  26%|██▌       | 17/66 [00:00&lt;00:02, 19.93it/s]
-Fetching 66 files:  47%|████▋     | 31/66 [00:01&lt;00:00, 39.55it/s]
-Fetching 66 files:  61%|██████    | 40/66 [00:01&lt;00:00, 29.77it/s]
-Fetching 66 files:  77%|███████▋  | 51/66 [00:01&lt;00:00, 38.15it/s]
-Fetching 66 files:  89%|████████▉ | 59/66 [00:01&lt;00:00, 42.51it/s]
-Fetching 66 files:  98%|█████████▊| 65/66 [00:01&lt;00:00, 44.76it/s]
-Fetching 66 files: 100%|██████████| 66/66 [00:01&lt;00:00, 35.09it/s]</div>
+Fetching 66 files:   2%|▏         | 1/66 [00:00&lt;00:11,  5.87it/s]
+Fetching 66 files:   3%|▎         | 2/66 [00:00&lt;00:10,  6.31it/s]
+Fetching 66 files:  17%|█▋        | 11/66 [00:00&lt;00:01, 34.40it/s]
+Fetching 66 files:  24%|██▍       | 16/66 [00:00&lt;00:01, 38.36it/s]
+Fetching 66 files:  32%|███▏      | 21/66 [00:00&lt;00:01, 23.86it/s]
+Fetching 66 files:  52%|█████▏    | 34/66 [00:01&lt;00:00, 39.79it/s]
+Fetching 66 files:  59%|█████▉    | 39/66 [00:01&lt;00:00, 28.73it/s]
+Fetching 66 files:  82%|████████▏ | 54/66 [00:01&lt;00:00, 40.50it/s]
+Fetching 66 files:  94%|█████████▍| 62/66 [00:01&lt;00:00, 45.88it/s]
+Fetching 66 files: 100%|██████████| 66/66 [00:01&lt;00:00, 37.06it/s]</div>
 <div class="cell-artifacts">
 <h4>Artifacts:</h4>
 <a href="artifacts/megablocks_run/megablocks_results.json" class="artifact" target="_blank">megablocks_results.json</a>
@@ -6668,7 +6669,7 @@ Fetching 66 files: 100%|██████████| 66/66 [00:01&lt;00:00, 3
 <span onclick="toggleOutput('visualization')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-visualization" onclick="toggleUvLogsFromHeader('visualization')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: visualization | deps: matplotlib | 3.14s
+Cell: visualization | deps: matplotlib | 3.13s
  | <button class="run-btn" onclick="runCell('visualization')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('visualization')">Copy</button>
 <a href="cells/visualization.py" target="_blank" class="raw-btn">Raw</a>
@@ -6914,30 +6915,30 @@ Loaded /repo/moe_benchmarks/megablocks_yamoe/.uvnote/cache/0febdf3420999533bc2e1
 Performance Summary:
 Implementation                 Avg (ms)     P95 (ms)     Tokens/sec   Relative Speed 
 --------------------------------------------------------------------------------
-megablocks_results                 3.87        8.54       25871        1.00x
-yamoe_results                      4.25        4.27       23536        0.91x
-binned_results                    37.04       39.33        2700        0.10x
-gptoss_training_results           46.92       51.42        2131        0.08x
-gptoss_results                    47.38       51.06        2111        0.08x
-
-Fastest: megablocks_results (3.87ms avg)
-Slowest: gptoss_results (47.38ms avg)
-Max Speedup: 12.3x
+megablocks_results                 3.85        8.50       25989        1.00x
+yamoe_results                      4.25        4.27       23549        0.91x
+binned_results                    35.79       37.38        2794        0.11x
+gptoss_results                    45.93       49.06        2177        0.08x
+gptoss_training_results           45.95       51.04        2176        0.08x
+
+Fastest: megablocks_results (3.85ms avg)
+Slowest: gptoss_training_results (45.95ms avg)
+Max Speedup: 11.9x
 </div>
 <div class="uv-install-logs" id="uv-logs-visualization">
 <div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
 <div class="uv-logs-content" style="display: none;">
+Downloading fonttools (4.7MiB)
 Downloading numpy (16.2MiB)
+Downloading matplotlib (8.3MiB)
 Downloading kiwisolver (1.4MiB)
 Downloading pillow (6.3MiB)
-Downloading fonttools (4.7MiB)
-Downloading matplotlib (8.3MiB)
  Downloading kiwisolver
  Downloading pillow
  Downloading fonttools
  Downloading matplotlib
  Downloading numpy
-Installed 11 packages in 48ms
+Installed 11 packages in 46ms
 </div>
 </div>
 <div class="cell-artifacts">