YMRohit Codex commited on
Commit
f77ebf7
·
1 Parent(s): 641764b

Submission media kit (video script, social, HF post) + README three-baseline accuracy

Browse files

Demo materials matched to the shipped UI (Build/Expert/Leaderboard/Run locally); hero number
sourced from the verified leaderboard, the recorded mint used to show the referee loop. README
now states both Local and Pro time vs eager, torch.compile, and max-autotune.

Co-authored-by: Codex <codex@openai.com>

CODEX_BRIEF_JUN14.md ADDED
@@ -0,0 +1,116 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Codex brief — submission media kit + README baseline accuracy (Jun 14)
2
+
3
+ You are working inside the live Hugging Face Space repo for **OUROBOROS Kernel Mint**
4
+ (`build-small-hackathon/ouroboros-kernel-mint`), a Build Small hackathon entry. Do exactly the
5
+ tasks below, then make ONE git commit with the trailer specified at the end. Do not touch any
6
+ file not listed. Do not run the app. Do not change `app.py`, `requirements.txt`, `style.css`,
7
+ `referee/`, or `replays/`.
8
+
9
+ ## Hard house style (non-negotiable)
10
+ - **No em-dashes anywhere.** Use commas, periods, "to", or "vs" instead.
11
+ - **No emoji in the new `submission/` files.** (In README, keep the existing emoji that are
12
+ already on the Links bullets, match the current style; do not add new ones.)
13
+ - Plain, confident, concrete voice. No marketing fluff, no "delve/leverage/seamless".
14
+ - **Honesty rule (critical): never claim the instant replay shows a 2x speedup.** The replays
15
+ are honest, modest wins over `torch.compile` (about 1.05x). The big numbers (around 2x) live
16
+ on the **Leaderboard** and in the verified corpus. The replay's job in the demo is to show the
17
+ full referee loop instantly, not to flex the biggest number.
18
+ - Only use the numbers given in the FACTS block. Do not invent or round differently.
19
+
20
+ ## FACTS (verified — use these verbatim, do not invent others)
21
+ - Space URL: `https://huggingface.co/spaces/build-small-hackathon/ouroboros-kernel-mint`
22
+ - Live UI: four tabs named **Build**, **Expert**, **Leaderboard**, **Run locally**.
23
+ - Build tab primary button: **"Mint kernel"**. Expert tab button: **"Mint from brief"**.
24
+ - Instant-demo button label (exact): **"Watch a recorded mint"**.
25
+ - Mode toggle: **"Local (offline)"** vs **Pro**. Local runs a 1B GGUF via llama.cpp on the
26
+ Space's own GPU and verifies in-process, no Modal/cloud calls. Pro runs the 27B on Modal.
27
+ - The in-app copy already says results are "timed against torch.compile and max-autotune (the
28
+ honest baselines), not just eager." Match that framing.
29
+ - Instant replays show, vs `torch.compile`: softmax ~1.05x, RMSNorm to gelu ~1.04x,
30
+ residual + LayerNorm to silu ~1.08x (each is also large vs unfused eager, but most of that gap
31
+ is just fusion the compiler also gets).
32
+ - Leaderboard hero numbers (all referee-verified, live on the board): `OUROBOROS-27B` champion
33
+ rows at softmax **2.04x**, softmax_scale **1.69x**, rmsnorm_tanh **1.56x**, rmsnorm_gelu
34
+ **1.54x** vs torch.compile; a top user row at softmax **2.10x**.
35
+ - Verified corpus headline: **76** model-written kernels beat `torch.compile` max-autotune on an
36
+ H200; **69** held up across 5 fresh re-benchmark runs (mean of means **1.30x**, range
37
+ **1.11x to 2.04x**); the other 7 are single-shot probes on never-trained problems. On a
38
+ **376-cell** grid of shapes and dtypes the trained kernels keep a **1.49x geomean** vs
39
+ max-autotune recompiled per cell, with about **10%** loss cells reported per cell.
40
+ - Models: **OpenBMB MiniCPM5-1B** (the default smith, genuinely 1B) and **Qwen3.6-27B** (Pro).
41
+ Both fine-tuned with the same loop: SFT on verified kernels, then RL where the only reward is
42
+ the referee's verdict. No human labels. Trained and served on **Modal** H200 (RL peaks ~110 GB
43
+ VRAM). The front end is one **Gradio** Space whose interactive surface is custom JavaScript, so
44
+ it is off-brand (not stock Gradio).
45
+ - Thesis: **the referee is the product** — the scarce thing is not the big model, it is a
46
+ verifier the model cannot fool.
47
+
48
+ ## Task 1 — create `submission/video-script.md`
49
+ A demo video script, target 2:00 to 2:15, that MATCHES the shipped UI above. Use this beat
50
+ sheet, fleshed out with tight voiceover and on-screen directions:
51
+
52
+ - 0:00–0:10 Cold open on the **Leaderboard** tab showing the `OUROBOROS-27B` 2.04x softmax
53
+ champion row. VO states the claim: small models wrote these kernels and a referee that cannot
54
+ be fooled verified every number.
55
+ - 0:10–0:30 **Build** tab: snap a normalization plus an activation together, the number-flow viz
56
+ animates. Explain what a fused GPU kernel is in one plain sentence.
57
+ - 0:30–0:55 Click **"Watch a recorded mint"**. Show the attempts stream, show one attempt FAIL,
58
+ then a pass. Walk the referee phases: compile, correctness vs PyTorch on adversarial inputs,
59
+ timing vs eager / torch.compile / max-autotune, verdict. Expand the real Triton source for a
60
+ beat. Say plainly that a green tick is earned, not asserted.
61
+ - 0:55–1:15 Do one **real live mint** (press "Mint kernel"); let the honest wait show; resolve
62
+ green. Note this one is live on the real backend.
63
+ - 1:15–1:40 Back to **Leaderboard** and the README receipts: the 76 / 69 / 1.30x / 1.49x numbers
64
+ and the honest bound (scheduling wins on memory-bound ops, about 10% of grid cells lose, shown
65
+ per cell).
66
+ - 1:40–2:00 Thesis and close on the Space URL.
67
+ - Production notes: 1920x1080, captions on (judges skim muted, put key numbers as on-screen
68
+ text), low driving music, cut on each verdict reveal, the replay covers a slow live mint but
69
+ still show one real mint, upload to YouTube unlisted and attach to the Space.
70
+ - Include a one-line warning: the voiceover must never state a number the screen contradicts; the
71
+ 2x hero number comes from the Leaderboard, the live mint is honestly whatever it measures.
72
+
73
+ ## Task 2 — create `submission/social-post.md`
74
+ Finalized posts ready to paste. Keep the strong true hook ("a 1B model beat PyTorch's compiler
75
+ 76 times, and the referee that scored it cannot be fooled"). Include:
76
+ - An X/Twitter primary post (under 280 chars per the main hook tweet, plus an optional 2-tweet
77
+ thread) tagging OpenBMB and Modal, with `[SPACE_URL]` and `[VIDEO_URL]` tokens.
78
+ - A longer LinkedIn version with the honest-bound sentence included.
79
+ - A short note reminding to paste the live post URL back into the README on submission.
80
+ Do not claim the replay shows 2x; the "76 times" and "beat the compiler" framing is the corpus
81
+ result and is fine.
82
+
83
+ ## Task 3 — create `submission/hf-post.md`
84
+ A ready-to-publish Hugging Face Post (about 120 to 160 words) for the field notes / fieldnotes
85
+ achievement and for distribution. Lead with the thesis, state the 76 / 1.30x / 1.49x numbers and
86
+ the honest bound in one line, link the Space and the corpus dataset
87
+ (`https://huggingface.co/datasets/YMRohit/ouroboros-kernel-corpus`), and end with an invitation
88
+ to mint one. Plain text, at most one or two tasteful emoji.
89
+
90
+ ## Task 4 — edit `README.md` (small, surgical)
91
+ 1. In "How to play" and "How it's built", make sure BOTH Local (offline) and Pro modes are
92
+ described as timing the kernel against **PyTorch eager, `torch.compile`, and `torch.compile`
93
+ max-autotune**. Read the current wording first; if a mode is described as eager-only or as
94
+ missing max-autotune, fix it to the three-baseline description. Match the in-app copy.
95
+ 2. Change the two Links bullets that currently read
96
+ `🎬 Demo video: *added with the submission*` and `📣 Social post: *added with the submission*`
97
+ to point at the in-repo media kit and make the final paste step obvious, for example:
98
+ `🎬 Demo video: shoot plan in [submission/video-script.md](submission/video-script.md); recording linked here on submission.`
99
+ `📣 Social post: text in [submission/social-post.md](submission/social-post.md) and [submission/hf-post.md](submission/hf-post.md); live link added here on submission.`
100
+ Keep it honest: do not assert a video or post URL that does not exist yet.
101
+
102
+ ## Final step — commit
103
+ Stage only the files you created or changed and commit once:
104
+
105
+ ```
106
+ git add submission/video-script.md submission/social-post.md submission/hf-post.md README.md CODEX_BRIEF_JUN14.md
107
+ git commit -m "Submission media kit (video script, social, HF post) + README three-baseline accuracy
108
+
109
+ Demo materials matched to the shipped UI (Build/Expert/Leaderboard/Run locally); hero number
110
+ sourced from the verified leaderboard, the recorded mint used to show the referee loop. README
111
+ now states both Local and Pro time vs eager, torch.compile, and max-autotune.
112
+
113
+ Co-authored-by: Codex <codex@openai.com>"
114
+ ```
115
+
116
+ Print `git log -1 --format='%H%n%an <%ae>%n%b'` at the end so the trailer can be verified.
README.md CHANGED
@@ -48,7 +48,8 @@ counts.
48
  few attempts. Each one faces the referee: compile, check correctness against PyTorch, then
49
  time it against PyTorch eager, `torch.compile`, and `torch.compile` max-autotune. Attempts
50
  that fail are shown, not hidden. Two ways to run it: **Local (offline)** does the whole loop
51
- inside this Space with no cloud calls, or **Pro** mints with the 27B on Modal.
 
52
  3. **Beat the compiler.** If your kernel is correct and faster than `torch.compile`
53
  max-autotune, it goes on the board. The crowns up top were minted by the 27B model. Switch
54
  on Pro mode to mint with that bigger model yourself.
@@ -115,8 +116,8 @@ not new algorithms.
115
  - 🤗 The 27B smith: [YMRohit/ouroboros-kernelsmith-qwen3.6-27b](https://huggingface.co/YMRohit/ouroboros-kernelsmith-qwen3.6-27b)
116
  - 🤗 Verified kernel corpus + evidence reports: [YMRohit/ouroboros-kernel-corpus](https://huggingface.co/datasets/YMRohit/ouroboros-kernel-corpus)
117
  - 📝 Field notes (what we built and learned): [FIELD_NOTES.md](https://huggingface.co/datasets/YMRohit/ouroboros-kernel-corpus/blob/main/FIELD_NOTES.md)
118
- - 🎬 Demo video: *added with the submission*
119
- - 📣 Social post: *added with the submission*
120
 
121
  Built for the Hugging Face Build Small hackathon (Backyard AI track). MIT licensed. Making
122
  small models fast is a problem for anyone who runs them locally, which is most of the reason I
 
48
  few attempts. Each one faces the referee: compile, check correctness against PyTorch, then
49
  time it against PyTorch eager, `torch.compile`, and `torch.compile` max-autotune. Attempts
50
  that fail are shown, not hidden. Two ways to run it: **Local (offline)** does the whole loop
51
+ inside this Space with no cloud calls, and **Pro** mints with the 27B on Modal. Both modes use
52
+ the same three-baseline comparison.
53
  3. **Beat the compiler.** If your kernel is correct and faster than `torch.compile`
54
  max-autotune, it goes on the board. The crowns up top were minted by the 27B model. Switch
55
  on Pro mode to mint with that bigger model yourself.
 
116
  - 🤗 The 27B smith: [YMRohit/ouroboros-kernelsmith-qwen3.6-27b](https://huggingface.co/YMRohit/ouroboros-kernelsmith-qwen3.6-27b)
117
  - 🤗 Verified kernel corpus + evidence reports: [YMRohit/ouroboros-kernel-corpus](https://huggingface.co/datasets/YMRohit/ouroboros-kernel-corpus)
118
  - 📝 Field notes (what we built and learned): [FIELD_NOTES.md](https://huggingface.co/datasets/YMRohit/ouroboros-kernel-corpus/blob/main/FIELD_NOTES.md)
119
+ - 🎬 Demo video: shoot plan in [submission/video-script.md](submission/video-script.md); recording linked here on submission.
120
+ - 📣 Social post: text in [submission/social-post.md](submission/social-post.md) and [submission/hf-post.md](submission/hf-post.md); live link added here on submission.
121
 
122
  Built for the Hugging Face Build Small hackathon (Backyard AI track). MIT licensed. Making
123
  small models fast is a problem for anyone who runs them locally, which is most of the reason I
submission/hf-post.md ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ The referee is the product.
2
+
3
+ OUROBOROS Kernel Mint is a Build Small hackathon Space where a 1B model writes real Triton GPU
4
+ kernels, then a verifier compiles them, checks correctness against PyTorch on adversarial inputs,
5
+ and times them against honest compiler baselines.
6
+
7
+ The verified corpus has 76 model-written kernels that beat `torch.compile` max-autotune on an
8
+ H200. 69 held up across 5 fresh re-benchmark runs, with a mean of means of 1.30x. On a 376-cell
9
+ grid of shapes and dtypes, trained kernels keep a 1.49x geomean against max-autotune recompiled
10
+ per cell. The honest bound: these are scheduling wins on memory-bound ops, about 10% of grid
11
+ cells lose, and the losses are reported per cell.
12
+
13
+ Space:
14
+ https://huggingface.co/spaces/build-small-hackathon/ouroboros-kernel-mint
15
+
16
+ Corpus:
17
+ https://huggingface.co/datasets/YMRohit/ouroboros-kernel-corpus
18
+
19
+ Mint one and see the referee loop.
submission/social-post.md ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Social Post
2
+
3
+ ## X/Twitter Primary
4
+
5
+ A 1B model beat PyTorch's compiler 76 times, and the referee that scored it cannot be fooled.
6
+
7
+ OUROBOROS Kernel Mint, built with @OpenBMB and @modal_labs.
8
+
9
+ Try it: [SPACE_URL]
10
+ Demo: [VIDEO_URL]
11
+
12
+ ## Optional X/Twitter Thread
13
+
14
+ First tweet:
15
+ A 1B model beat PyTorch's compiler 76 times, and the referee that scored it cannot be fooled.
16
+ OUROBOROS Kernel Mint turns block-built GPU ops into real Triton kernels, then verifies them
17
+ against PyTorch and compiler baselines. [SPACE_URL]
18
+
19
+ Second tweet:
20
+ The instant replay is for the referee loop: compile, correctness, timing, verdict. The big
21
+ numbers live on the Leaderboard and verified corpus. Built with OpenBMB models and Modal H200s.
22
+ Demo: [VIDEO_URL]
23
+
24
+ ## LinkedIn
25
+
26
+ A 1B model beat PyTorch's compiler 76 times, and the referee that scored it cannot be fooled.
27
+
28
+ OUROBOROS Kernel Mint is a Build Small hackathon Space where you build a GPU operation out of
29
+ blocks, then a fine-tuned OpenBMB MiniCPM5-1B writes a real Triton kernel for it. A separate
30
+ referee compiles the kernel, checks correctness against PyTorch on adversarial inputs, and times
31
+ it against PyTorch eager, `torch.compile`, and `torch.compile` max-autotune.
32
+
33
+ The thesis is simple: the referee is the product. The scarce thing is not the big model, it is a
34
+ verifier the model cannot fool.
35
+
36
+ The verified corpus has 76 model-written kernels that beat `torch.compile` max-autotune on an
37
+ H200, with 69 holding up across 5 fresh re-benchmark runs. The honest bound: these are scheduling
38
+ wins on memory-bound ops, about 10% of grid cells lose, and the losses are reported per cell.
39
+
40
+ Built with OpenBMB models and Modal H200s.
41
+
42
+ Space: [SPACE_URL]
43
+ Demo: [VIDEO_URL]
44
+
45
+ ## Submission Reminder
46
+
47
+ After posting, paste the live social post URL back into the README Links section.
submission/video-script.md ADDED
@@ -0,0 +1,76 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Demo Video Script
2
+
3
+ Target runtime: 2:00 to 2:15.
4
+
5
+ ## 0:00 to 0:10, cold open
6
+
7
+ On screen: Open the **Leaderboard** tab. Frame the `OUROBOROS-27B` softmax champion row at
8
+ **2.04x** vs `torch.compile`.
9
+
10
+ Voiceover: "Small models wrote these kernels, and a referee that cannot be fooled verified every
11
+ number. This is OUROBOROS Kernel Mint."
12
+
13
+ On-screen text: `2.04x softmax, referee-verified`
14
+
15
+ ## 0:10 to 0:30, Build
16
+
17
+ On screen: Switch to **Build**. Snap a normalization block and an activation together. Let the
18
+ number-flow visualization animate.
19
+
20
+ Voiceover: "A fused GPU kernel combines several neural-network steps into one small program, so
21
+ the GPU moves data fewer times. Here the machine is built out of blocks, then the model has to
22
+ write the Triton code."
23
+
24
+ ## 0:30 to 0:55, recorded referee loop
25
+
26
+ On screen: Click **"Watch a recorded mint"**. Show the attempts stream. Hold briefly on one
27
+ attempt marked FAIL, then cut to a passing attempt. Expand the real Triton source for a beat.
28
+
29
+ Voiceover: "The replay shows the full referee loop instantly. Each attempt is compiled, checked
30
+ for correctness against PyTorch on adversarial inputs, timed against eager, `torch.compile`, and
31
+ `torch.compile` max-autotune, then given a verdict. A green tick is earned, not asserted."
32
+
33
+ On-screen text: `compile -> correctness -> timing -> verdict`
34
+
35
+ ## 0:55 to 1:15, live mint
36
+
37
+ On screen: Return to **Build** and press **"Mint kernel"**. Let the wait show. When the result
38
+ lands, resolve on the green state.
39
+
40
+ Voiceover: "Now this is a real live mint on the backend. The model writes a kernel, the same
41
+ referee checks it, and the measured result is whatever it earns."
42
+
43
+ On-screen text: `live backend, real verdict`
44
+
45
+ ## 1:15 to 1:40, receipts
46
+
47
+ On screen: Go back to **Leaderboard**, then show the README receipts.
48
+
49
+ Voiceover: "The corpus has **76** model-written kernels that beat `torch.compile` max-autotune on
50
+ an H200. **69** held up across 5 fresh re-benchmark runs, with a mean of means of **1.30x**. On a
51
+ **376-cell** grid of shapes and dtypes, trained kernels keep a **1.49x geomean** against
52
+ max-autotune recompiled per cell. The honest bound is visible too: these are scheduling wins on
53
+ memory-bound ops, and about **10%** of grid cells lose, reported per cell."
54
+
55
+ On-screen text: `76 wins, 69 rechecked, 1.30x mean, 1.49x grid geomean`
56
+
57
+ ## 1:40 to 2:00, thesis and close
58
+
59
+ On screen: End on the Space URL:
60
+ `https://huggingface.co/spaces/build-small-hackathon/ouroboros-kernel-mint`
61
+
62
+ Voiceover: "The referee is the product. The scarce thing is not the big model. It is a verifier
63
+ the model cannot fool. Mint one yourself."
64
+
65
+ On-screen text: `the referee is the product`
66
+
67
+ ## Production Notes
68
+
69
+ - Record at 1920x1080.
70
+ - Captions on. Judges skim muted, so put key numbers as on-screen text.
71
+ - Use low driving music.
72
+ - Cut on each verdict reveal.
73
+ - The replay covers a slow live mint, but still show one real mint.
74
+ - Upload to YouTube unlisted and attach it to the Space.
75
+ - Warning: the voiceover must never state a number the screen contradicts. The 2x hero number
76
+ comes from the Leaderboard, while the live mint is honestly whatever it measures.