YMRohit/icml2026-62070-foam-logbook-artifacts
5.71 MB
Rohit Yelukati Mahendra PRO
YMRohit
AI & ML interests
None yet
Recent Activity
updated a Space 32 minutes ago
YMRohit/icml2026-62070-foam-logbook published a Space 32 minutes ago
YMRohit/icml2026-62070-foam-logbook updated a bucket 32 minutes ago
YMRohit/icml2026-62070-foam-logbook-artifactsOrganizations
Posts 1
Post
3060
A 1B model that writes GPU kernels you can trust
I fine-tuned OpenBMB's MiniCPM5-1B to write Triton GPU kernels, then let an immutable referee decide if they are real: compile, check correctness against PyTorch on adversarial inputs, time against eager, torch.compile, and torch.compile max-autotune, then block the known ways of gaming the benchmark.
The 1B setup beat torch.compile max-autotune in 12/12 independently seeded runs. The larger Qwen3.6-27B smith pushed the same referee loop further: 76 verified compiler-beating kernels on H200, with 69 surviving a 5-run stability gate and 7 kept as single-shot probes on unseen problems. On a 376-cell shape/dtype grid, the stability-gated kernels keep a 1.49x geomean, with about 10% of cells losing and reported per cell.
Honest bound: these are scheduling wins on memory-bound ops, not new algorithms or wins over cuBLAS/FlashAttention. The scarce thing is not the big model, it is the verifier it cannot fool.
Full write-up: https://huggingface.co/blog/YMRohit/ouroboros-kernel-mint
Try it: build-small-hackathon/ouroboros-kernel-mint
2-min demo: https://youtu.be/ViicZHktb-A
Built for #BuildSmallHackathon with MiniCPM, Qwen, Triton, Gradio, Codex, and Modal H200s.
I fine-tuned OpenBMB's MiniCPM5-1B to write Triton GPU kernels, then let an immutable referee decide if they are real: compile, check correctness against PyTorch on adversarial inputs, time against eager, torch.compile, and torch.compile max-autotune, then block the known ways of gaming the benchmark.
The 1B setup beat torch.compile max-autotune in 12/12 independently seeded runs. The larger Qwen3.6-27B smith pushed the same referee loop further: 76 verified compiler-beating kernels on H200, with 69 surviving a 5-run stability gate and 7 kept as single-shot probes on unseen problems. On a 376-cell shape/dtype grid, the stability-gated kernels keep a 1.49x geomean, with about 10% of cells losing and reported per cell.
Honest bound: these are scheduling wins on memory-bound ops, not new algorithms or wins over cuBLAS/FlashAttention. The scarce thing is not the big model, it is the verifier it cannot fool.
Full write-up: https://huggingface.co/blog/YMRohit/ouroboros-kernel-mint
Try it: build-small-hackathon/ouroboros-kernel-mint
2-min demo: https://youtu.be/ViicZHktb-A
Built for #BuildSmallHackathon with MiniCPM, Qwen, Triton, Gradio, Codex, and Modal H200s.
Logbooks, results, and artifacts for reproductions of ICML 2026 papers 61269 (d3LLM) and 64918 (MemoryBench).
-
Repro - d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation
🎯Browse and share experiment logs with a coding agent
-
YMRohit/icml2026-61269-d3llm-repro-results
Updated • 83 -
Repro - MemoryBench: A Benchmark for Memory and Continual Learning in LLM Systems
🎯Explore benchmark logs and sync findings with an AI agent
-
YMRohit/icml2026-64918-memorybench-repro-results
Preview • Updated • 80
Logbooks, results, and artifacts for reproductions of ICML 2026 papers 61269 (d3LLM) and 64918 (MemoryBench).
-
Repro - d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation
🎯Browse and share experiment logs with a coding agent
-
YMRohit/icml2026-61269-d3llm-repro-results
Updated • 83 -
Repro - MemoryBench: A Benchmark for Memory and Continual Learning in LLM Systems
🎯Explore benchmark logs and sync findings with an AI agent
-
YMRohit/icml2026-64918-memorybench-repro-results
Preview • Updated • 80
Independent theory/numerical audit, GPU Jobs, reviewer-adjudicated verdicts, Trackio logbook, full artifact bundle, and poster.
spaces 16
Running
Repro - FOAM: Blocked State Folding for Memory-Efficient LLM Training
🧪
Explore experiment logs and collaborate with an AI agent
Running
Repro - MemoryBench: A Benchmark for Memory and Continual Learning in LLM Systems
🎯
Explore benchmark logs and sync findings with an AI agent
Running
Agents
Icml64918 Memorybench
🎯
Show interactive tracking dashboard
Running
Reproducing Maximum Likelihood Reinforcement Learning (ICML 2026 #65332)
🧮
Browse experiment logbook and collaborate with a coding agent
Running
Repro - Optimal Unconstrained Self-Distillation in Ridge Regression
🔬
Explore and collaborate on project logbooks online
Running
Repro - d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation
🎯
Browse and share experiment logs with a coding agent
models 16
YMRohit/ouroboros-kernelsmith-minicpm5-1b
Text Generation • Updated • 9
YMRohit/ouroboros-kernelsmith-minicpm5-1b-GGUF
Text Generation • 1B • Updated • 87 • 1
YMRohit/ouroboros-kernelsmith-qwen3.6-27b
Text Generation • Updated
YMRohit/rohit-portfolio-lfm25-350m-onnx
Text Generation • Updated • 4
YMRohit/smollm-135m-funcalling-rl-gpo
Updated
YMRohit/qwen0.5reasoning
Updated
YMRohit/GRPO-latest-checkpoint
Updated
YMRohit/sqltest
Updated
YMRohit/Llama-3.2-11B-Vision-Instruct-Radiology
Updated
YMRohit/Qwen2-VL-7B-Instruct-Radiology-Mini
Updated
datasets 6
YMRohit/icml2026-64918-memorybench-repro-results
Preview • Updated • 80
YMRohit/icml2026-61269-d3llm-repro-results
Updated • 83
YMRohit/icml20531-adam-degenerate-repro
Updated • 49
YMRohit/icml17708-grok-grokking-repro
Updated • 66
YMRohit/ouroboros-kernel-corpus
Viewer • Updated • 2.46k • 128 • 1
YMRohit/EmotionReasoning
Viewer • Updated • 20k • 104