Tracking Prioritization of eager operations

[jmalone-tt]

🎯 Updated Optimal Operation Prioritization Strategy

After filtering out removed operations, we now have 141 unique operations across 156 models.

🥇 TIER 1: Ultra-Critical (1 operation)

aten.convolution.default - 128 models (82.1%) - The core convolution operation

🥈 TIER 2: High-Value Foundation (3 operations)

1. aten.view.default - 119 models (76.3%) - Tensor reshaping
2. aten.add.Tensor - 114 models (73.1%) - Element-wise addition
3. aten.addmm.default - 96 models (61.5%) - Matrix multiplication + bias

🏗️ TIER 3: Foundation Operations (5 operations)

4. aten.relu.default - 87 models (55.8%) - ReLU activation
5. aten.mean.dim - 85 models (54.5%) - Mean along dimensions
6. aten.cat.default - 68 models (43.6%) - Tensor concatenation
7. aten._native_batch_norm_legit_no_training.default - 65 models (41.7%) - Batch normalization
8. aten.mul.Tensor - 64 models (41.0%) - Element-wise multiplication

🚀 Quick Wins Strategy (16 operations)

Target the simplest 10 models first (2-6 operations each):

• Autoencoder (linear) - 2 ops
• Autoencoder (conv) - 3 ops
• VGG models (11/13/16/19) - 5 ops each
• U-Net variants - 5-6 ops

Phase 1 operations needed for these simple models:

1. aten.addmm.default
2. aten.relu.default
3. aten.convolution.default
4. aten.max_pool2d_with_indices.default
5. aten.view.default
6. aten._native_batch_norm_legit_no_training.default
7. Plus 10 more specialized operations listed in per-model documentation

🎯 Recommended Implementation Path:

1. Start with aten.convolution.default (82% model coverage foundation)
2. Add Tier 2 operations (view, add, addmm) for broad tensor manipulation support
3. Implement Phase 1 set (16 ops) to unlock 10 complete simple models
4. Complete Tier 3 for 40-60% model coverage
5. Follow incremental strategy by operation count levels