NVFP4 first 3 problems release
#11
by
marksaroufim
- opened
- .gitignore +0 -1
- README.md +7 -66
- docs.md +0 -158
- nvidia_nvfp4_submissions.parquet +0 -3
- queries.sql +0 -124
- scripts/nvfp4/analyze_submissions.py +0 -168
- scripts/nvfp4/get_fastest_submission.py +0 -20
- scripts/nvfp4/query_submissions.py +0 -57
.gitignore
DELETED
|
@@ -1 +0,0 @@
|
|
| 1 |
-
__pycache__/
|
|
|
|
|
|
README.md
CHANGED
|
@@ -1,86 +1,27 @@
|
|
| 1 |
---
|
| 2 |
configs:
|
| 3 |
-
- config_name:
|
| 4 |
data_files: "submissions.parquet"
|
| 5 |
-
- config_name:
|
| 6 |
data_files: "successful_submissions.parquet"
|
| 7 |
-
- config_name: nvidia_nvfp4_submissions
|
| 8 |
-
data_files: "nvidia_nvfp4_submissions.parquet"
|
| 9 |
- config_name: leaderboards
|
| 10 |
data_files: "leaderboards.parquet"
|
| 11 |
tags:
|
| 12 |
- code
|
| 13 |
-
license:
|
| 14 |
---
|
| 15 |
|
| 16 |
-
|
| 17 |
|
| 18 |
-
|
| 19 |
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
### AMD MI300 Submissions
|
| 23 |
-
| File | Description |
|
| 24 |
-
|------|-------------|
|
| 25 |
-
| `submissions.parquet` | All AMD competition submissions |
|
| 26 |
-
| `successful_submissions.parquet` | AMD submissions that passed correctness tests |
|
| 27 |
-
| `deduplicated_submissions.parquet` | AMD submissions deduplicated by (user, code) |
|
| 28 |
-
| `deduplicated_successful_submissions.parquet` | Deduplicated passing AMD submissions |
|
| 29 |
-
|
| 30 |
-
**AMD Problems:** fp8-gemm, moe (mixture of experts), mla-decode, all2all, gemm+reducescatter, allgather+gemm
|
| 31 |
-
|
| 32 |
-
### NVIDIA Blackwell NVFP4 Submissions
|
| 33 |
-
| File | Size | Description |
|
| 34 |
-
|------|------|-------------|
|
| 35 |
-
| `nvidia_nvfp4_submissions.parquet` | ~1.4 GB | NVFP4 submissions deduplicated by (user, code), with full code content |
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
**NVFP4 Problems:** gemv (leaderboard 595), gemm (597), dual_gemm (598), modal_dual_gemm (697)
|
| 39 |
-
|
| 40 |
-
**Note on Dual GEMM:** There are two variants of the dual_gemm problem. Midway through the competition, on-prem hardware measurements became unreliable, so a second leaderboard was created on Modal infrastructure. The Modal measurements (leaderboard 697, `modal_nvfp4_dual_gemm`) are more trustworthy.
|
| 41 |
-
|
| 42 |
-
**Note:** Scores are execution time in seconds. **Lower is better.**
|
| 43 |
-
|
| 44 |
-
## Helper Scripts
|
| 45 |
-
|
| 46 |
-
- `analyze_submissions.py` - Python functions for analyzing submissions
|
| 47 |
-
- `skills.md` - Documentation for data processing workflows
|
| 48 |
-
|
| 49 |
-
### Quick Start
|
| 50 |
-
|
| 51 |
-
```python
|
| 52 |
-
from analyze_submissions import load_submissions, top_contestants, author_progression
|
| 53 |
-
|
| 54 |
-
# Load NVIDIA NVFP4 data
|
| 55 |
-
df = load_submissions()
|
| 56 |
-
|
| 57 |
-
# Get top 20 for a problem
|
| 58 |
-
leaders = top_contestants(df, problem_name='nvfp4_gemm', n=20)
|
| 59 |
-
|
| 60 |
-
# See a user's progression over time
|
| 61 |
-
progression = author_progression(df, user_name='username', problem_name='nvfp4_gemm')
|
| 62 |
-
```
|
| 63 |
-
|
| 64 |
-
## Learn More
|
| 65 |
-
|
| 66 |
-
- Competition platform: [gpumode.com](https://gpumode.com)
|
| 67 |
-
- Reference kernels and problem specs: [github.com/gpu-mode/reference-kernels](https://github.com/gpu-mode/reference-kernels)
|
| 68 |
-
|
| 69 |
-
## License
|
| 70 |
-
|
| 71 |
-
This dataset is licensed under the [Creative Commons Attribution 4.0 International License (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/).
|
| 72 |
-
|
| 73 |
-
You are free to share and adapt the material for any purpose, even commercially, provided you give appropriate credit.
|
| 74 |
-
|
| 75 |
-
**Attribution:** Please cite GPU Mode and link to this dataset. For academic papers, use the citation below.
|
| 76 |
-
|
| 77 |
-
## Citation
|
| 78 |
|
| 79 |
If you use this dataset in your work, please cite:
|
| 80 |
|
| 81 |
```bibtex
|
| 82 |
@inproceedings{
|
| 83 |
-
|
| 84 |
title={KernelBot: A Competition Platform for Writing Heterogeneous {GPU} Code},
|
| 85 |
author={Alex L Zhang and Matej Sirovatka and Erik Schultheis and Benjamin Horowitz and Mark Saroufim},
|
| 86 |
booktitle={Championing Open-source DEvelopment in ML Workshop @ ICML25},
|
|
|
|
| 1 |
---
|
| 2 |
configs:
|
| 3 |
+
- config_name: submissions
|
| 4 |
data_files: "submissions.parquet"
|
| 5 |
+
- config_name: successful_submissions
|
| 6 |
data_files: "successful_submissions.parquet"
|
|
|
|
|
|
|
| 7 |
- config_name: leaderboards
|
| 8 |
data_files: "leaderboards.parquet"
|
| 9 |
tags:
|
| 10 |
- code
|
| 11 |
+
license: mit
|
| 12 |
---
|
| 13 |
|
| 14 |
+
This is the dataset that was created from the first and second AMD $100K kernel competitions, containing roughly 110K kernels for fp8-gemm, moe, mla, all2all, gemm+reducescatter, and allgather+gemm optimized to run on MI300. Learn more at gpumode.com/v2/news
|
| 15 |
|
| 16 |
+
To see the full list of kernel competitions we've ran and are running you can checkout https://github.com/gpu-mode/reference-kernels which also contains details on reference kernels and their input shapes and distributions
|
| 17 |
|
| 18 |
+
We are planning on adding kernels optimized for NVFP4 on Blackwell next
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
|
| 20 |
If you use this dataset in your work, please cite:
|
| 21 |
|
| 22 |
```bibtex
|
| 23 |
@inproceedings{
|
| 24 |
+
zhang2025kernelbot,
|
| 25 |
title={KernelBot: A Competition Platform for Writing Heterogeneous {GPU} Code},
|
| 26 |
author={Alex L Zhang and Matej Sirovatka and Erik Schultheis and Benjamin Horowitz and Mark Saroufim},
|
| 27 |
booktitle={Championing Open-source DEvelopment in ML Workshop @ ICML25},
|
docs.md
DELETED
|
@@ -1,158 +0,0 @@
|
|
| 1 |
-
# Kernelbot Data Processing Skills
|
| 2 |
-
|
| 3 |
-
This document describes how to extract and process submission data from the Kernelbot database.
|
| 4 |
-
|
| 5 |
-
## Database Connection
|
| 6 |
-
|
| 7 |
-
The production database is hosted on Heroku. **NEVER run write operations (INSERT, UPDATE, DELETE) on this database.**
|
| 8 |
-
|
| 9 |
-
```bash
|
| 10 |
-
# Get DATABASE_URL from Heroku
|
| 11 |
-
heroku config:get DATABASE_URL --app discord-cluster-manager
|
| 12 |
-
```
|
| 13 |
-
|
| 14 |
-
## Database Schema
|
| 15 |
-
|
| 16 |
-
The relevant tables are in the `leaderboard` schema:
|
| 17 |
-
|
| 18 |
-
| Table | Description |
|
| 19 |
-
|-------|-------------|
|
| 20 |
-
| `leaderboard.leaderboard` | Problem definitions (id, name, deadline, task, description) |
|
| 21 |
-
| `leaderboard.submission` | User submissions (id, leaderboard_id, user_id, code_id, submission_time, status) |
|
| 22 |
-
| `leaderboard.runs` | Execution results (submission_id, score, passed, mode, runner, result) |
|
| 23 |
-
| `leaderboard.user_info` | User details (id, user_name) |
|
| 24 |
-
| `leaderboard.gpu_type` | GPU types per problem (leaderboard_id, gpu_type) |
|
| 25 |
-
| `leaderboard.code_files` | Actual submission code content (old_code text, code bytea) |
|
| 26 |
-
|
| 27 |
-
## Key Problem IDs
|
| 28 |
-
|
| 29 |
-
### NVFP4 Problems
|
| 30 |
-
- **595**: nvfp4_gemv
|
| 31 |
-
- **597**: nvfp4_gemm
|
| 32 |
-
- **598**: nvfp4_dual_gemm
|
| 33 |
-
- **730**: nvfp4_group_gemm (not released yet)
|
| 34 |
-
|
| 35 |
-
### AMD Problems
|
| 36 |
-
- **398**: amd-identity
|
| 37 |
-
- **399**: amd-fp8-mm
|
| 38 |
-
- **430**: amd-mixture-of-experts
|
| 39 |
-
- **463**: amd-mla-decode
|
| 40 |
-
- **563**: amd-all2all
|
| 41 |
-
- **564**: amd-gemm-rs
|
| 42 |
-
- **565**: amd-ag-gemm
|
| 43 |
-
|
| 44 |
-
## Run Modes
|
| 45 |
-
|
| 46 |
-
| Mode | Description | Has Score? |
|
| 47 |
-
|------|-------------|------------|
|
| 48 |
-
| `test` | Correctness tests | No |
|
| 49 |
-
| `benchmark` | Performance benchmarks (internal) | No |
|
| 50 |
-
| `leaderboard` | Official leaderboard runs | **Yes** |
|
| 51 |
-
| `profile.0-3` | Profiling runs | No |
|
| 52 |
-
|
| 53 |
-
**Important:**
|
| 54 |
-
- Use `mode = 'leaderboard'` when joining runs to get scores.
|
| 55 |
-
- **Lower scores are better** (scores are execution time in seconds).
|
| 56 |
-
|
| 57 |
-
## SQL Queries
|
| 58 |
-
|
| 59 |
-
All SQL queries are in `queries.sql`. Key queries:
|
| 60 |
-
- List all problems
|
| 61 |
-
- Check submission counts
|
| 62 |
-
- Export deduplicated submissions with code
|
| 63 |
-
- Get top N submissions
|
| 64 |
-
- Get user progression over time
|
| 65 |
-
|
| 66 |
-
## Adding Support for a New Problem
|
| 67 |
-
|
| 68 |
-
### Step 1: Find the Problem ID
|
| 69 |
-
Use the "LIST ALL PROBLEMS" query from `queries.sql`.
|
| 70 |
-
|
| 71 |
-
### Step 2: Check Submission Counts
|
| 72 |
-
Use the "CHECK SUBMISSION COUNTS" query from `queries.sql`.
|
| 73 |
-
|
| 74 |
-
### Step 3: Export Deduplicated Submissions
|
| 75 |
-
Use the "EXPORT DEDUPLICATED SUBMISSIONS WITH CODE" query from `queries.sql`.
|
| 76 |
-
|
| 77 |
-
```python
|
| 78 |
-
import pandas as pd
|
| 79 |
-
import psycopg2
|
| 80 |
-
|
| 81 |
-
DATABASE_URL = "..." # from heroku config:get
|
| 82 |
-
conn = psycopg2.connect(DATABASE_URL)
|
| 83 |
-
|
| 84 |
-
# Read query from queries.sql and modify problem IDs as needed
|
| 85 |
-
with open('queries.sql') as f:
|
| 86 |
-
# Find and use the export query section
|
| 87 |
-
pass
|
| 88 |
-
|
| 89 |
-
df = pd.read_sql(query, conn)
|
| 90 |
-
df.to_parquet('new_problem_submissions.parquet', index=False)
|
| 91 |
-
```
|
| 92 |
-
|
| 93 |
-
### Step 4: Verify Data Quality
|
| 94 |
-
```python
|
| 95 |
-
from analyze_submissions import load_submissions, leaderboard_summary
|
| 96 |
-
|
| 97 |
-
df = load_submissions('new_problem_submissions.parquet')
|
| 98 |
-
print(leaderboard_summary(df))
|
| 99 |
-
```
|
| 100 |
-
|
| 101 |
-
## Accessing Submission Code
|
| 102 |
-
|
| 103 |
-
The parquet files include the full code content for each submission:
|
| 104 |
-
|
| 105 |
-
```python
|
| 106 |
-
from analyze_submissions import load_submissions
|
| 107 |
-
|
| 108 |
-
df = load_submissions()
|
| 109 |
-
|
| 110 |
-
# Get a specific user's best submission
|
| 111 |
-
user_subs = df[(df['user_name'] == 'gau.nernst') & (df['problem_name'] == 'nvfp4_gemv')]
|
| 112 |
-
best = user_subs.sort_values('score').head(1)
|
| 113 |
-
|
| 114 |
-
# Access the code
|
| 115 |
-
code = best['code'].values[0]
|
| 116 |
-
print(code)
|
| 117 |
-
```
|
| 118 |
-
|
| 119 |
-
## Helper Functions
|
| 120 |
-
|
| 121 |
-
Use `analyze_submissions.py`:
|
| 122 |
-
|
| 123 |
-
```python
|
| 124 |
-
from analyze_submissions import (
|
| 125 |
-
load_submissions, # Load parquet file
|
| 126 |
-
author_progression, # See user's submissions over time
|
| 127 |
-
top_contestants, # Get leaderboard rankings
|
| 128 |
-
leaderboard_summary, # Summary stats per problem
|
| 129 |
-
user_stats, # Stats for a specific user
|
| 130 |
-
format_score # Format score with units (us, ms, s)
|
| 131 |
-
)
|
| 132 |
-
```
|
| 133 |
-
|
| 134 |
-
## Environment Setup
|
| 135 |
-
|
| 136 |
-
```bash
|
| 137 |
-
uv venv .venv
|
| 138 |
-
source .venv/bin/activate
|
| 139 |
-
uv pip install pandas pyarrow psycopg2-binary
|
| 140 |
-
```
|
| 141 |
-
|
| 142 |
-
## Files
|
| 143 |
-
|
| 144 |
-
| File | Description |
|
| 145 |
-
|------|-------------|
|
| 146 |
-
| `nvidia_nvfp4_submissions.parquet` | Deduplicated NVIDIA NVFP4 submissions with code (~1.4 GB) |
|
| 147 |
-
| `queries.sql` | All SQL queries for data extraction |
|
| 148 |
-
| `scripts/nvfp4/analyze_submissions.py` | Helper functions library |
|
| 149 |
-
| `scripts/nvfp4/get_fastest_submission.py` | Print user's fastest submission |
|
| 150 |
-
| `scripts/nvfp4/query_submissions.py` | List submission IDs or query specific ID |
|
| 151 |
-
|
| 152 |
-
## Review Checklist Before Pushing
|
| 153 |
-
|
| 154 |
-
1. Verify submission counts match expectations
|
| 155 |
-
2. Check for any anomalies in scores (negative, extremely large, etc.)
|
| 156 |
-
3. Confirm deduplication worked correctly
|
| 157 |
-
4. Test helper functions work with the new data
|
| 158 |
-
5. Run `python scripts/nvfp4/query_submissions.py` to verify
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
nvidia_nvfp4_submissions.parquet
DELETED
|
@@ -1,3 +0,0 @@
|
|
| 1 |
-
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:64f4d29e68f3471b524be55a62f1af58a4d36a0f1f2d191c59a8f5302941e664
|
| 3 |
-
size 1622580981
|
|
|
|
|
|
|
|
|
|
|
|
queries.sql
DELETED
|
@@ -1,124 +0,0 @@
|
|
| 1 |
-
-- Kernelbot Database Queries
|
| 2 |
-
-- All queries are READ ONLY. Never run INSERT/UPDATE/DELETE on production.
|
| 3 |
-
-- Scores are execution time in seconds. Lower is better.
|
| 4 |
-
|
| 5 |
-
--------------------------------------------------------------------------------
|
| 6 |
-
-- LIST ALL PROBLEMS
|
| 7 |
-
--------------------------------------------------------------------------------
|
| 8 |
-
SELECT
|
| 9 |
-
l.id,
|
| 10 |
-
l.name,
|
| 11 |
-
l.deadline,
|
| 12 |
-
l.description,
|
| 13 |
-
array_agg(g.gpu_type) as gpu_types
|
| 14 |
-
FROM leaderboard.leaderboard l
|
| 15 |
-
LEFT JOIN leaderboard.gpu_type g ON l.id = g.leaderboard_id
|
| 16 |
-
GROUP BY l.id, l.name, l.deadline, l.description
|
| 17 |
-
ORDER BY l.id;
|
| 18 |
-
|
| 19 |
-
--------------------------------------------------------------------------------
|
| 20 |
-
-- PROBLEM IDS
|
| 21 |
-
--------------------------------------------------------------------------------
|
| 22 |
-
-- NVFP4: 595 (gemv), 597 (gemm), 598 (dual_gemm), 730 (group_gemm)
|
| 23 |
-
-- AMD: 398 (identity), 399 (fp8-mm), 430 (moe), 463 (mla-decode),
|
| 24 |
-
-- 563 (all2all), 564 (gemm-rs), 565 (ag-gemm)
|
| 25 |
-
|
| 26 |
-
--------------------------------------------------------------------------------
|
| 27 |
-
-- CHECK SUBMISSION COUNTS FOR A PROBLEM
|
| 28 |
-
--------------------------------------------------------------------------------
|
| 29 |
-
SELECT
|
| 30 |
-
COUNT(*) as total_submissions,
|
| 31 |
-
COUNT(DISTINCT user_id) as unique_users
|
| 32 |
-
FROM leaderboard.submission
|
| 33 |
-
WHERE leaderboard_id = 595; -- Replace with problem ID
|
| 34 |
-
|
| 35 |
-
--------------------------------------------------------------------------------
|
| 36 |
-
-- EXPORT DEDUPLICATED SUBMISSIONS WITH CODE
|
| 37 |
-
-- Deduplicates by (user_id, code_id), keeping the fastest score
|
| 38 |
-
--------------------------------------------------------------------------------
|
| 39 |
-
WITH ranked AS (
|
| 40 |
-
SELECT
|
| 41 |
-
s.id as submission_id,
|
| 42 |
-
s.leaderboard_id,
|
| 43 |
-
l.name as problem_name,
|
| 44 |
-
s.user_id,
|
| 45 |
-
u.user_name,
|
| 46 |
-
s.code_id,
|
| 47 |
-
s.file_name,
|
| 48 |
-
s.submission_time,
|
| 49 |
-
s.status,
|
| 50 |
-
r.score,
|
| 51 |
-
r.passed,
|
| 52 |
-
r.mode,
|
| 53 |
-
r.runner,
|
| 54 |
-
COALESCE(c.old_code, convert_from(c.code, 'UTF8')) as code,
|
| 55 |
-
ROW_NUMBER() OVER (
|
| 56 |
-
PARTITION BY s.leaderboard_id, s.user_id, s.code_id
|
| 57 |
-
ORDER BY r.score ASC NULLS LAST
|
| 58 |
-
) as rn
|
| 59 |
-
FROM leaderboard.submission s
|
| 60 |
-
JOIN leaderboard.leaderboard l ON s.leaderboard_id = l.id
|
| 61 |
-
LEFT JOIN leaderboard.user_info u ON s.user_id = u.id
|
| 62 |
-
LEFT JOIN leaderboard.runs r ON s.id = r.submission_id AND r.mode = 'leaderboard'
|
| 63 |
-
LEFT JOIN leaderboard.code_files c ON s.code_id = c.id
|
| 64 |
-
WHERE s.leaderboard_id IN (595, 597, 598) -- Replace with problem IDs
|
| 65 |
-
)
|
| 66 |
-
SELECT
|
| 67 |
-
submission_id, leaderboard_id, problem_name, user_id, user_name,
|
| 68 |
-
code_id, file_name, submission_time, status, score, passed, mode, runner, code
|
| 69 |
-
FROM ranked
|
| 70 |
-
WHERE rn = 1
|
| 71 |
-
ORDER BY problem_name, score ASC NULLS LAST;
|
| 72 |
-
|
| 73 |
-
--------------------------------------------------------------------------------
|
| 74 |
-
-- CHECK RUN MODES AND SCORES
|
| 75 |
-
--------------------------------------------------------------------------------
|
| 76 |
-
SELECT
|
| 77 |
-
r.mode,
|
| 78 |
-
COUNT(*) as cnt,
|
| 79 |
-
COUNT(r.score) as has_score,
|
| 80 |
-
MIN(r.score) as min_score,
|
| 81 |
-
MAX(r.score) as max_score
|
| 82 |
-
FROM leaderboard.runs r
|
| 83 |
-
JOIN leaderboard.submission s ON r.submission_id = s.id
|
| 84 |
-
WHERE s.leaderboard_id IN (595, 597, 598)
|
| 85 |
-
GROUP BY r.mode
|
| 86 |
-
ORDER BY cnt DESC;
|
| 87 |
-
|
| 88 |
-
--------------------------------------------------------------------------------
|
| 89 |
-
-- GET TOP N SUBMISSIONS FOR A PROBLEM
|
| 90 |
-
--------------------------------------------------------------------------------
|
| 91 |
-
SELECT
|
| 92 |
-
u.user_name,
|
| 93 |
-
r.score,
|
| 94 |
-
s.submission_time
|
| 95 |
-
FROM leaderboard.submission s
|
| 96 |
-
JOIN leaderboard.runs r ON s.id = r.submission_id AND r.mode = 'leaderboard'
|
| 97 |
-
LEFT JOIN leaderboard.user_info u ON s.user_id = u.id
|
| 98 |
-
WHERE s.leaderboard_id = 595 -- Replace with problem ID
|
| 99 |
-
AND r.passed = true
|
| 100 |
-
AND r.score IS NOT NULL
|
| 101 |
-
ORDER BY r.score ASC
|
| 102 |
-
LIMIT 20;
|
| 103 |
-
|
| 104 |
-
--------------------------------------------------------------------------------
|
| 105 |
-
-- GET USER'S SUBMISSIONS OVER TIME (progression)
|
| 106 |
-
--------------------------------------------------------------------------------
|
| 107 |
-
SELECT
|
| 108 |
-
s.submission_time,
|
| 109 |
-
r.score,
|
| 110 |
-
r.passed
|
| 111 |
-
FROM leaderboard.submission s
|
| 112 |
-
JOIN leaderboard.runs r ON s.id = r.submission_id AND r.mode = 'leaderboard'
|
| 113 |
-
JOIN leaderboard.user_info u ON s.user_id = u.id
|
| 114 |
-
WHERE u.user_name = 'gau.nernst' -- Replace with username
|
| 115 |
-
AND s.leaderboard_id = 595 -- Replace with problem ID
|
| 116 |
-
ORDER BY s.submission_time ASC;
|
| 117 |
-
|
| 118 |
-
--------------------------------------------------------------------------------
|
| 119 |
-
-- GET CODE FOR A SPECIFIC SUBMISSION
|
| 120 |
-
--------------------------------------------------------------------------------
|
| 121 |
-
SELECT
|
| 122 |
-
COALESCE(c.old_code, convert_from(c.code, 'UTF8')) as code
|
| 123 |
-
FROM leaderboard.code_files c
|
| 124 |
-
WHERE c.id = 79741; -- Replace with code_id
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
scripts/nvfp4/analyze_submissions.py
DELETED
|
@@ -1,168 +0,0 @@
|
|
| 1 |
-
#!/usr/bin/env python3
|
| 2 |
-
"""
|
| 3 |
-
Helper functions for analyzing kernelbot submissions.
|
| 4 |
-
|
| 5 |
-
Usage:
|
| 6 |
-
from analyze_submissions import load_submissions, author_progression, top_contestants
|
| 7 |
-
"""
|
| 8 |
-
|
| 9 |
-
import pandas as pd
|
| 10 |
-
from pathlib import Path
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
def format_score(score, unit='us'):
|
| 14 |
-
"""
|
| 15 |
-
Format score with appropriate units.
|
| 16 |
-
|
| 17 |
-
Args:
|
| 18 |
-
score: Score in seconds
|
| 19 |
-
unit: 'us' for microseconds, 'ms' for milliseconds, 'auto' for automatic
|
| 20 |
-
|
| 21 |
-
Returns:
|
| 22 |
-
Formatted string with units
|
| 23 |
-
"""
|
| 24 |
-
if pd.isna(score):
|
| 25 |
-
return 'N/A'
|
| 26 |
-
|
| 27 |
-
if unit == 'auto':
|
| 28 |
-
if score < 0.001: # Less than 1ms, show in microseconds
|
| 29 |
-
return f"{score * 1_000_000:.2f} µs"
|
| 30 |
-
elif score < 1: # Less than 1s, show in milliseconds
|
| 31 |
-
return f"{score * 1_000:.3f} ms"
|
| 32 |
-
else:
|
| 33 |
-
return f"{score:.4f} s"
|
| 34 |
-
elif unit == 'us':
|
| 35 |
-
return f"{score * 1_000_000:.2f} µs"
|
| 36 |
-
elif unit == 'ms':
|
| 37 |
-
return f"{score * 1_000:.3f} ms"
|
| 38 |
-
else:
|
| 39 |
-
return f"{score:.6f} s"
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
def load_submissions(parquet_path: str = None) -> pd.DataFrame:
|
| 43 |
-
"""Load deduplicated submissions from parquet file."""
|
| 44 |
-
if parquet_path is None:
|
| 45 |
-
parquet_path = Path(__file__).parent.parent.parent / "nvidia_nvfp4_submissions.parquet"
|
| 46 |
-
return pd.read_parquet(parquet_path)
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
def author_progression(df: pd.DataFrame, user_id: str = None, user_name: str = None,
|
| 50 |
-
problem_name: str = None) -> pd.DataFrame:
|
| 51 |
-
"""
|
| 52 |
-
Get submissions from an author sorted by time to see their progression.
|
| 53 |
-
|
| 54 |
-
Args:
|
| 55 |
-
df: DataFrame of submissions
|
| 56 |
-
user_id: Filter by user ID (Discord ID)
|
| 57 |
-
user_name: Filter by username (partial match, case-insensitive)
|
| 58 |
-
problem_name: Filter by problem name
|
| 59 |
-
|
| 60 |
-
Returns:
|
| 61 |
-
DataFrame sorted by submission_time showing the author's journey
|
| 62 |
-
"""
|
| 63 |
-
result = df.copy()
|
| 64 |
-
|
| 65 |
-
if user_id:
|
| 66 |
-
result = result[result['user_id'] == user_id]
|
| 67 |
-
|
| 68 |
-
if user_name:
|
| 69 |
-
result = result[result['user_name'].str.contains(user_name, case=False, na=False)]
|
| 70 |
-
|
| 71 |
-
if problem_name:
|
| 72 |
-
result = result[result['problem_name'] == problem_name]
|
| 73 |
-
|
| 74 |
-
return result.sort_values('submission_time')
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
def top_contestants(df: pd.DataFrame, problem_name: str = None, n: int = 20,
|
| 78 |
-
passing_only: bool = True) -> pd.DataFrame:
|
| 79 |
-
"""
|
| 80 |
-
Get top contestants sorted by their best score (fastest time).
|
| 81 |
-
|
| 82 |
-
Args:
|
| 83 |
-
df: DataFrame of submissions
|
| 84 |
-
problem_name: Filter by problem name (required for meaningful results)
|
| 85 |
-
n: Number of top contestants to return
|
| 86 |
-
passing_only: Only include passing submissions
|
| 87 |
-
|
| 88 |
-
Returns:
|
| 89 |
-
DataFrame with top contestants and their best scores
|
| 90 |
-
"""
|
| 91 |
-
result = df.copy()
|
| 92 |
-
|
| 93 |
-
if problem_name:
|
| 94 |
-
result = result[result['problem_name'] == problem_name]
|
| 95 |
-
|
| 96 |
-
if passing_only:
|
| 97 |
-
result = result[result['passed'] == True]
|
| 98 |
-
|
| 99 |
-
# Filter out rows with NA scores
|
| 100 |
-
result = result.dropna(subset=['score'])
|
| 101 |
-
|
| 102 |
-
if result.empty:
|
| 103 |
-
return pd.DataFrame(columns=['user_name', 'user_id', 'score', 'submission_time', 'problem_name'])
|
| 104 |
-
|
| 105 |
-
# Get best score per user
|
| 106 |
-
best_scores = result.loc[result.groupby('user_id')['score'].idxmin()]
|
| 107 |
-
|
| 108 |
-
return best_scores.sort_values('score').head(n)[
|
| 109 |
-
['user_name', 'user_id', 'score', 'submission_time', 'problem_name']
|
| 110 |
-
]
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
def leaderboard_summary(df: pd.DataFrame, score_unit='us') -> pd.DataFrame:
|
| 114 |
-
"""
|
| 115 |
-
Get summary statistics for each problem.
|
| 116 |
-
|
| 117 |
-
Args:
|
| 118 |
-
df: DataFrame of submissions
|
| 119 |
-
score_unit: 'us' for microseconds, 'ms' for milliseconds, 's' for seconds
|
| 120 |
-
|
| 121 |
-
Returns:
|
| 122 |
-
DataFrame with submission counts, unique users, score ranges
|
| 123 |
-
"""
|
| 124 |
-
summary = df.groupby('problem_name').agg({
|
| 125 |
-
'submission_id': 'count',
|
| 126 |
-
'user_id': 'nunique',
|
| 127 |
-
'score': ['min', 'median', 'max'],
|
| 128 |
-
'passed': 'sum'
|
| 129 |
-
})
|
| 130 |
-
|
| 131 |
-
summary.columns = ['submissions', 'unique_users', 'best_score', 'median_score',
|
| 132 |
-
'worst_score', 'passing_count']
|
| 133 |
-
|
| 134 |
-
# Convert scores to specified unit
|
| 135 |
-
if score_unit == 'us':
|
| 136 |
-
multiplier = 1_000_000
|
| 137 |
-
summary['best_score'] = (summary['best_score'] * multiplier).round(2)
|
| 138 |
-
summary['median_score'] = (summary['median_score'] * multiplier).round(2)
|
| 139 |
-
summary['worst_score'] = (summary['worst_score'] * multiplier).round(2)
|
| 140 |
-
elif score_unit == 'ms':
|
| 141 |
-
multiplier = 1_000
|
| 142 |
-
summary['best_score'] = (summary['best_score'] * multiplier).round(3)
|
| 143 |
-
summary['median_score'] = (summary['median_score'] * multiplier).round(3)
|
| 144 |
-
summary['worst_score'] = (summary['worst_score'] * multiplier).round(3)
|
| 145 |
-
|
| 146 |
-
return summary
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
def user_stats(df: pd.DataFrame, user_id: str = None, user_name: str = None) -> pd.DataFrame:
|
| 150 |
-
"""
|
| 151 |
-
Get statistics for a specific user across all problems.
|
| 152 |
-
"""
|
| 153 |
-
result = df.copy()
|
| 154 |
-
|
| 155 |
-
if user_id:
|
| 156 |
-
result = result[result['user_id'] == user_id]
|
| 157 |
-
elif user_name:
|
| 158 |
-
result = result[result['user_name'].str.contains(user_name, case=False, na=False)]
|
| 159 |
-
|
| 160 |
-
return result.groupby('problem_name').agg({
|
| 161 |
-
'submission_id': 'count',
|
| 162 |
-
'score': 'min',
|
| 163 |
-
'passed': 'sum'
|
| 164 |
-
}).rename(columns={
|
| 165 |
-
'submission_id': 'num_submissions',
|
| 166 |
-
'score': 'best_score',
|
| 167 |
-
'passed': 'passing_count'
|
| 168 |
-
})
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
scripts/nvfp4/get_fastest_submission.py
DELETED
|
@@ -1,20 +0,0 @@
|
|
| 1 |
-
#!/usr/bin/env python3
|
| 2 |
-
"""Print gau.nernst's fastest submission code to stdout."""
|
| 3 |
-
|
| 4 |
-
import pandas as pd
|
| 5 |
-
from pathlib import Path
|
| 6 |
-
|
| 7 |
-
df = pd.read_parquet(Path(__file__).parent.parent.parent / 'nvidia_nvfp4_submissions.parquet')
|
| 8 |
-
|
| 9 |
-
# Get fastest submission across all problems
|
| 10 |
-
best = df[df['user_name'] == 'gau.nernst'].sort_values('score').head(1)
|
| 11 |
-
|
| 12 |
-
problem = best['problem_name'].values[0]
|
| 13 |
-
score_us = best['score'].values[0] * 1_000_000
|
| 14 |
-
|
| 15 |
-
print(f"User: gau.nernst")
|
| 16 |
-
print(f"Problem: {problem}")
|
| 17 |
-
print(f"Score: {score_us:.2f} µs")
|
| 18 |
-
print(f"Submission ID: {best['submission_id'].values[0]}")
|
| 19 |
-
print("\n=== CODE ===\n")
|
| 20 |
-
print(best['code'].values[0])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
scripts/nvfp4/query_submissions.py
DELETED
|
@@ -1,57 +0,0 @@
|
|
| 1 |
-
#!/usr/bin/env python3
|
| 2 |
-
"""
|
| 3 |
-
Query submissions by user/problem or by submission ID.
|
| 4 |
-
|
| 5 |
-
Usage:
|
| 6 |
-
python query_submissions.py # Show all submission IDs for gau.nernst on gemv
|
| 7 |
-
python query_submissions.py --id 187476 # Show code for specific submission ID
|
| 8 |
-
python query_submissions.py --user gau.nernst --problem nvfp4_gemm
|
| 9 |
-
"""
|
| 10 |
-
|
| 11 |
-
import argparse
|
| 12 |
-
import pandas as pd
|
| 13 |
-
from pathlib import Path
|
| 14 |
-
|
| 15 |
-
df = pd.read_parquet(Path(__file__).parent.parent.parent / 'nvidia_nvfp4_submissions.parquet')
|
| 16 |
-
|
| 17 |
-
parser = argparse.ArgumentParser()
|
| 18 |
-
parser.add_argument('--id', type=int, help='Submission ID to query')
|
| 19 |
-
parser.add_argument('--user', default='gau.nernst', help='Username to filter')
|
| 20 |
-
parser.add_argument('--problem', default='nvfp4_gemv', help='Problem name to filter')
|
| 21 |
-
args = parser.parse_args()
|
| 22 |
-
|
| 23 |
-
if args.id:
|
| 24 |
-
# Query specific submission
|
| 25 |
-
sub = df[df['submission_id'] == args.id]
|
| 26 |
-
if sub.empty:
|
| 27 |
-
print(f"Submission {args.id} not found")
|
| 28 |
-
else:
|
| 29 |
-
row = sub.iloc[0]
|
| 30 |
-
score_us = row['score'] * 1_000_000 if pd.notna(row['score']) else 'N/A'
|
| 31 |
-
print(f"ID: {row['submission_id']}")
|
| 32 |
-
print(f"User: {row['user_name']}")
|
| 33 |
-
print(f"Problem: {row['problem_name']}")
|
| 34 |
-
print(f"Score: {score_us:.2f} µs" if isinstance(score_us, float) else f"Score: {score_us}")
|
| 35 |
-
print(f"\n=== CODE ===\n")
|
| 36 |
-
print(row['code'])
|
| 37 |
-
else:
|
| 38 |
-
# List all submission IDs for user/problem
|
| 39 |
-
subs = df[(df['user_name'] == args.user) & (df['problem_name'] == args.problem)]
|
| 40 |
-
subs = subs.sort_values('score')
|
| 41 |
-
|
| 42 |
-
ids = subs['submission_id'].tolist()
|
| 43 |
-
scores = [(row['submission_id'], row['score'] * 1_000_000 if pd.notna(row['score']) else None)
|
| 44 |
-
for _, row in subs.iterrows()]
|
| 45 |
-
|
| 46 |
-
print(f"User: {args.user} | Problem: {args.problem} | Count: {len(ids)}")
|
| 47 |
-
print(f"\nSubmission IDs (sorted by score, fastest first):")
|
| 48 |
-
print(ids)
|
| 49 |
-
|
| 50 |
-
# Get fastest/slowest with valid scores
|
| 51 |
-
valid_scores = [(sid, sc) for sid, sc in scores if sc is not None]
|
| 52 |
-
if valid_scores:
|
| 53 |
-
print(f"\nFastest: {valid_scores[0][0]} ({valid_scores[0][1]:.2f} µs)")
|
| 54 |
-
print(f"Slowest: {valid_scores[-1][0]} ({valid_scores[-1][1]:.2f} µs)")
|
| 55 |
-
print(f"\nQuery a specific submission: python query_submissions.py --id {valid_scores[0][0]}")
|
| 56 |
-
else:
|
| 57 |
-
print("\nNo submissions with scores found")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|