sleepy
8e72eef09c
- Rename gamma to glm5 and model to minimax-m2.7 - Add model_comparison/ directory with head-to-head analyses - Sanitize all session.jsonl files: remove absolute paths and usernames - Remove __pycache__ artifacts - Add .gitignore
21 lines
676 B
Markdown
21 lines
676 B
Markdown
Design and implement a high-performance fused softmax + top-k kernel in CUDA (or CUDA-like pseudocode).
|
|
|
|
Requirements:
|
|
- Input: logits [B, T, V]
|
|
- Output:
|
|
- top-k indices per (B, T)
|
|
- top-k probabilities (after softmax)
|
|
|
|
Constraints:
|
|
1. Do NOT materialize the full softmax matrix in global memory.
|
|
2. Must be numerically stable (log-sum-exp).
|
|
3. Minimize global memory reads/writes.
|
|
4. Use shared memory where appropriate.
|
|
5. Handle large V (e.g., 50k+) efficiently.
|
|
|
|
Deliver:
|
|
- Kernel pseudocode or CUDA code
|
|
- Memory access pattern explanation
|
|
- Warp-level optimization strategy
|
|
- Complexity analysis (bandwidth vs compute bound)
|
|
- Comparison to naive implementation |