YMRohit/icml27364-kernelbounds-bucket
3.66 MB
Rohit Yelukati Mahendra
YMRohit
AI & ML interests
None yet
Recent Activity
liked a model about 9 hours ago
nvidia/Nemotron-3-Diarization updated a collection 2 months ago
ICML 2026 reproductions — d3LLM & MemoryBench (Claude) updated a Space 2 months ago
YMRohit/icml2026-27364-kernel-bounds-logbookOrganizations
Posts 1
Post
3072
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 • 46 -
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 • 74
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 • 46 -
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 • 74
Independent theory/numerical audit, GPU Jobs, reviewer-adjudicated verdicts, Trackio logbook, full artifact bundle, and poster.
spaces 26
Running
Repro - On Uniform Error Bounds for Kernel Regression under Non-Gaussian Noise
🎯
Explore and edit a collaborative research logbook
Paused
Agents
Icml27364 Kernelbounds
🎯
Display experiment tracking results in an interactive view
Running
Repro - Softmax as Linear Attention in the Large-Prompt Regime: a Measure-based Perspective
🎯
Organize research notes and sync with an AI coding agent
Paused
Agents
Icml5886 Softmaxlinear
🎯
Show a visual timeline of your program's I/O activity
Running
Repro - Can Adaptive Gradient Methods Converge under Heavy-Tailed Noise? A Case Study of AdaGrad
🎯
Browse experiment logs and sync findings with an AI agent
Paused
Agents
Icml31042 Adagradheavy
🎯
Show an interactive tracking dashboard for your data
models 16
YMRohit/ouroboros-kernelsmith-minicpm5-1b
Text Generation • Updated • 9
YMRohit/ouroboros-kernelsmith-minicpm5-1b-GGUF
Text Generation • 1B • Updated • 59 • 1
YMRohit/ouroboros-kernelsmith-qwen3.6-27b
Text Generation • Updated
YMRohit/rohit-portfolio-lfm25-350m-onnx
Text Generation • Updated • 14
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 • 74
YMRohit/icml2026-61269-d3llm-repro-results
Updated • 46
YMRohit/icml20531-adam-degenerate-repro
Updated • 32
YMRohit/icml17708-grok-grokking-repro
Updated • 149
YMRohit/ouroboros-kernel-corpus
Viewer • Updated • 2.46k • 72 • 1
YMRohit/EmotionReasoning
Viewer • Updated • 20k • 43