Spaces:
Running on Zero
Running on Zero
Add bounded track/network GUI smoke test; no fallback or performance claim
Browse files- README.md +8 -2
- ZEROGPU_MANIFEST.json +4 -3
- app.py +21 -0
- training_smoke.py +89 -0
README.md
CHANGED
|
@@ -10,7 +10,7 @@ app_file: app.py
|
|
| 10 |
pinned: false
|
| 11 |
---
|
| 12 |
|
| 13 |
-
Private native CUDA compatibility probe, not a qualified trainer.
|
| 14 |
Uses included ZeroGPU allowance through a 45-second leased call. No paid Jobs
|
| 15 |
or dedicated hardware requested. Check the authenticated caller's included quota
|
| 16 |
before each call; the app itself does not know account billing balances.
|
|
@@ -22,6 +22,12 @@ to that revision and upstream d8ab74c25df70b6a86ed101080e33d5fb785ea6f.
|
|
| 22 |
|
| 23 |
No CUDA initialization outside the lease. First check device discovery and a
|
| 24 |
three-value sine kernel in process, then the normal Rust gpu-info subprocess.
|
| 25 |
-
No fallback,
|
| 26 |
must independently pass before testing training. Retained PACKAGE_MANIFEST.json
|
| 27 |
describes the prior Docker snapshot only; ZEROGPU_MANIFEST.json covers this overlay.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 10 |
pinned: false
|
| 11 |
---
|
| 12 |
|
| 13 |
+
Private native CUDA compatibility probe and bounded upload smoke test, not a qualified trainer.
|
| 14 |
Uses included ZeroGPU allowance through a 45-second leased call. No paid Jobs
|
| 15 |
or dedicated hardware requested. Check the authenticated caller's included quota
|
| 16 |
before each call; the app itself does not know account billing balances.
|
|
|
|
| 22 |
|
| 23 |
No CUDA initialization outside the lease. First check device discovery and a
|
| 24 |
three-value sine kernel in process, then the normal Rust gpu-info subprocess.
|
| 25 |
+
No fallback, trajectory parity, speedup or H200 claim. A second lease
|
| 26 |
must independently pass before testing training. Retained PACKAGE_MANIFEST.json
|
| 27 |
describes the prior Docker snapshot only; ZEROGPU_MANIFEST.json covers this overlay.
|
| 28 |
+
|
| 29 |
+
The upload test validates the exact supplied A01 track and a 6-4-2 Formula network
|
| 30 |
+
before leasing a GPU for at most 90 seconds. It repeats qualification, then runs
|
| 31 |
+
4 vehicles for 1 generation of 60 physics ticks, seed 42, proton math profile.
|
| 32 |
+
The subprocess has a 25-second timeout. It exports a candidate and receipt, not
|
| 33 |
+
an improvement claim. Real-game import compatibility and physics parity remain unverified.
|
ZEROGPU_MANIFEST.json
CHANGED
|
@@ -1,9 +1,10 @@
|
|
| 1 |
{
|
| 2 |
-
"README.md": "
|
| 3 |
-
"app.py": "
|
| 4 |
"probe_core.py": "97a36ad8cf8efec09d84bb4b2c1ee6794658ceedfd643cd9849ff3c6d85aca9d",
|
| 5 |
"requirements.txt": "d4a97517b799a94b3370ea2e11aee3e41304aa12b77bab9abecd0b37e961acaf",
|
| 6 |
"BINARY_PROVENANCE.json": "f2223c1c6e2bbf8f6101b46e11563f4d866f99094c529cd5584f3228309e6551",
|
| 7 |
"bin/altd-sim": "2f2f2144ae14b0ef599ca8510f95712da00243c6875881491d384b394e7499f7",
|
| 8 |
-
"bin/libaltd_gpu.so": "279d6353b4c691686d6aeeae6abfd4a82b959e9b604f079a6c45f9823b48a87a"
|
|
|
|
| 9 |
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"README.md": "cd5ad5b4e2dd5230ecc0eb3083ea40e90c3665b5a9529a4286fbead2ed37bacb",
|
| 3 |
+
"app.py": "5d7e2fe152b0846f73a8383ef9e87c60925b4f7384221a4b04f639c5bdeb1e81",
|
| 4 |
"probe_core.py": "97a36ad8cf8efec09d84bb4b2c1ee6794658ceedfd643cd9849ff3c6d85aca9d",
|
| 5 |
"requirements.txt": "d4a97517b799a94b3370ea2e11aee3e41304aa12b77bab9abecd0b37e961acaf",
|
| 6 |
"BINARY_PROVENANCE.json": "f2223c1c6e2bbf8f6101b46e11563f4d866f99094c529cd5584f3228309e6551",
|
| 7 |
"bin/altd-sim": "2f2f2144ae14b0ef599ca8510f95712da00243c6875881491d384b394e7499f7",
|
| 8 |
+
"bin/libaltd_gpu.so": "279d6353b4c691686d6aeeae6abfd4a82b959e9b604f079a6c45f9823b48a87a",
|
| 9 |
+
"training_smoke.py": "91e278b8fdf1a025d2ce455f5a4e55b1751691be2689fadda92955fc6dcbb508"
|
| 10 |
}
|
app.py
CHANGED
|
@@ -1,6 +1,7 @@
|
|
| 1 |
import spaces # Must precede libraries that interact with CUDA.
|
| 2 |
import gradio as gr
|
| 3 |
from probe_core import qualify, verify_artifacts
|
|
|
|
| 4 |
|
| 5 |
# Integrity only: no device initialization or CUDA library loading at startup.
|
| 6 |
verify_artifacts()
|
|
@@ -9,6 +10,14 @@ verify_artifacts()
|
|
| 9 |
def probe():
|
| 10 |
return qualify()
|
| 11 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
with gr.Blocks() as demo:
|
| 13 |
gr.Markdown("# TAO · ZeroGPU compatibility check\n"
|
| 14 |
"Private, bounded native-library qualification. No training, CPU fallback, "
|
|
@@ -17,6 +26,18 @@ with gr.Blocks() as demo:
|
|
| 17 |
button = gr.Button("Run one compatibility check")
|
| 18 |
result = gr.JSON(label="Qualification receipt")
|
| 19 |
button.click(probe, inputs=[], outputs=result, api_name="probe", concurrency_limit=1)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
|
| 21 |
if __name__ == "__main__":
|
| 22 |
demo.queue(max_size=1).launch()
|
|
|
|
| 1 |
import spaces # Must precede libraries that interact with CUDA.
|
| 2 |
import gradio as gr
|
| 3 |
from probe_core import qualify, verify_artifacts
|
| 4 |
+
from training_smoke import validate_uploads, execute_smoke
|
| 5 |
|
| 6 |
# Integrity only: no device initialization or CUDA library loading at startup.
|
| 7 |
verify_artifacts()
|
|
|
|
| 10 |
def probe():
|
| 11 |
return qualify()
|
| 12 |
|
| 13 |
+
@spaces.GPU(duration=90)
|
| 14 |
+
def leased_smoke(template):
|
| 15 |
+
return execute_smoke(template)
|
| 16 |
+
|
| 17 |
+
def smoke(track, network):
|
| 18 |
+
# Reject unsupported inputs on CPU before requesting a GPU lease.
|
| 19 |
+
return leased_smoke(validate_uploads(track, network))
|
| 20 |
+
|
| 21 |
with gr.Blocks() as demo:
|
| 22 |
gr.Markdown("# TAO · ZeroGPU compatibility check\n"
|
| 23 |
"Private, bounded native-library qualification. No training, CPU fallback, "
|
|
|
|
| 26 |
button = gr.Button("Run one compatibility check")
|
| 27 |
result = gr.JSON(label="Qualification receipt")
|
| 28 |
button.click(probe, inputs=[], outputs=result, api_name="probe", concurrency_limit=1)
|
| 29 |
+
gr.Markdown("## Bounded upload test\nUpload the supplied A01 track and a 6–4–2 Formula .evolution. "
|
| 30 |
+
"Fixed at 4 vehicles, 1 generation, 60 physics ticks, seed 42. "
|
| 31 |
+
"Checks compatibility again inside a new lease. Up to 90 seconds of included quota. "
|
| 32 |
+
"No CPU fallback, parity or performance claim. A candidate is not evidence of improvement.")
|
| 33 |
+
track = gr.File(label="A01 Beginner track", file_types=[".track"], type="filepath")
|
| 34 |
+
network = gr.File(label="Formula network", file_types=[".evolution"], type="filepath")
|
| 35 |
+
run = gr.Button("Run bounded training test")
|
| 36 |
+
report = gr.JSON(label="Training receipt")
|
| 37 |
+
candidate = gr.File(label="Candidate network")
|
| 38 |
+
receipt = gr.File(label="Download receipt")
|
| 39 |
+
run.click(smoke, inputs=[track, network], outputs=[report, candidate, receipt],
|
| 40 |
+
api_name="smoke", concurrency_limit=1)
|
| 41 |
|
| 42 |
if __name__ == "__main__":
|
| 43 |
demo.queue(max_size=1).launch()
|
training_smoke.py
ADDED
|
@@ -0,0 +1,89 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""One fixed, small GUI upload test. Not a general-purpose training endpoint."""
|
| 2 |
+
import gzip
|
| 3 |
+
import hashlib
|
| 4 |
+
import json
|
| 5 |
+
import os
|
| 6 |
+
from pathlib import Path
|
| 7 |
+
import subprocess
|
| 8 |
+
import tempfile
|
| 9 |
+
import time
|
| 10 |
+
|
| 11 |
+
ROOT = Path(__file__).parent
|
| 12 |
+
|
| 13 |
+
def validate_uploads(track, network):
|
| 14 |
+
if not track or not network:
|
| 15 |
+
raise ValueError('Upload both A01 Beginner.track and a Formula .evolution network')
|
| 16 |
+
track, network = Path(track), Path(network)
|
| 17 |
+
if track.stat().st_size > 1024*1024 or network.stat().st_size > 1024*1024:
|
| 18 |
+
raise ValueError('Smoke-test uploads must be at most 1 MB each')
|
| 19 |
+
expected = ROOT / 'tracks/A01 Beginner.track'
|
| 20 |
+
if hashlib.sha256(track.read_bytes()).digest() != hashlib.sha256(expected.read_bytes()).digest():
|
| 21 |
+
raise ValueError('This first smoke test accepts only the supplied A01 Beginner.track fixture')
|
| 22 |
+
# Reuse the deployed dashboard parser and all its shape/finite-value/sensor guards.
|
| 23 |
+
from server import parse_network
|
| 24 |
+
template, _ = parse_network(network.read_bytes())
|
| 25 |
+
if template['NeuralNetwork']['Shape'] != [6, 4, 2]:
|
| 26 |
+
raise ValueError('First smoke test requires a 6-4-2 Formula network')
|
| 27 |
+
return template
|
| 28 |
+
|
| 29 |
+
def execute_smoke(template):
|
| 30 |
+
from probe_core import qualify
|
| 31 |
+
qualification = qualify()
|
| 32 |
+
if not qualification['qualified_for_next_test']:
|
| 33 |
+
return {'status':'qualification_failed','qualification':qualification}, None, None
|
| 34 |
+
directory = Path(tempfile.mkdtemp(prefix='ald-smoke-'))
|
| 35 |
+
n = template['NeuralNetwork']
|
| 36 |
+
network = {'shape':n['Shape'],'weights':n['Weights'],'biases':n['Biases'],
|
| 37 |
+
'inputs':[f'sensor-{i}' for i in range(6)], 'outputs':['Steering','Acceleration']}
|
| 38 |
+
model = {'sensor_layout':{'names':network['inputs'],'sensors':template['Sensors']}}
|
| 39 |
+
settings = {'population':4,'selection_algorithm':'best','selection_size':1,
|
| 40 |
+
'crossover':'none','mutation_rate':0.01,'adaptive_mutation':False,'weight_decay':0,
|
| 41 |
+
'preserve_parents':'on_custom','preserve_parents_size':1,
|
| 42 |
+
'rewards':[{'metric':'total_score','type':'default','weight':1}]}
|
| 43 |
+
for name, data in [('network',network),('model',model),('settings',settings)]:
|
| 44 |
+
(directory/(name+'.json')).write_text(json.dumps(data))
|
| 45 |
+
binary = ROOT/'bin/altd-sim'
|
| 46 |
+
binary.chmod(0o755)
|
| 47 |
+
out = directory/'native'
|
| 48 |
+
cmd = [str(binary),'--math-profile','proton','--threads','1','train-scratch',
|
| 49 |
+
'--gpu','--scene',str(ROOT/'tracks/A01 Beginner.scene.json'),
|
| 50 |
+
'--spawn-trace',str(ROOT/'tracks/A01 Beginner.spawn.json'),
|
| 51 |
+
'--network',str(directory/'network.json'),'--init-network',str(directory/'network.json'),
|
| 52 |
+
'--model',str(directory/'model.json'),'--settings',str(directory/'settings.json'),
|
| 53 |
+
'--shape','6,4,2','--population','4','--generations','1','--ticks','60',
|
| 54 |
+
'--seed','42','--batch-count','2','--eliminate-on-wall','--idle-eliminate',
|
| 55 |
+
'--out-dir',str(out),'--mutation-start','0.01','--mutation-end','0.01',
|
| 56 |
+
'--schedule','linear','--checkpoint-every','1','--save-final-candidate']
|
| 57 |
+
start = time.perf_counter()
|
| 58 |
+
try:
|
| 59 |
+
result = subprocess.run(cmd, capture_output=True,text=True,timeout=25,
|
| 60 |
+
cwd=ROOT/'github-simulator/source',
|
| 61 |
+
env=dict(os.environ,ALTD_GPU_LIB=str(ROOT/'bin/libaltd_gpu.so')))
|
| 62 |
+
except subprocess.TimeoutExpired:
|
| 63 |
+
report = {'status':'timeout','qualification':qualification,'timeout_seconds':25,
|
| 64 |
+
'parity_verified':False,'speedup_verified':False}
|
| 65 |
+
receipt = directory/'receipt.json'
|
| 66 |
+
receipt.write_text(json.dumps(report,indent=2)+'\n')
|
| 67 |
+
return report, None, str(receipt)
|
| 68 |
+
report = {'status':'failed','qualification':qualification,'population':4,'generations':1,
|
| 69 |
+
'ticks':60,'seed':42,'math_profile':'proton','returncode':result.returncode,
|
| 70 |
+
'wall_seconds':time.perf_counter()-start,'stdout':result.stdout[-8192:],
|
| 71 |
+
'stderr':result.stderr[-8192:],'parity_verified':False,'speedup_verified':False,
|
| 72 |
+
'network_sha256':hashlib.sha256((directory/'network.json').read_bytes()).hexdigest(),
|
| 73 |
+
'scene_sha256':hashlib.sha256((ROOT/'tracks/A01 Beginner.scene.json').read_bytes()).hexdigest(),
|
| 74 |
+
'command':cmd}
|
| 75 |
+
candidate = None
|
| 76 |
+
if result.returncode == 0 and (out/'candidate.json').exists():
|
| 77 |
+
trained = json.loads((out/'candidate.json').read_text())
|
| 78 |
+
template['NeuralNetwork']['Weights'] = trained['weights']
|
| 79 |
+
template['NeuralNetwork']['Biases'] = trained['biases']
|
| 80 |
+
template['Name'] = 'ZeroGPU smoke candidate — not benchmarked'
|
| 81 |
+
candidate = directory/'candidate.evolution'
|
| 82 |
+
candidate.write_bytes(gzip.compress(json.dumps(template).encode(),mtime=0))
|
| 83 |
+
from server import parse_network
|
| 84 |
+
parse_network(candidate.read_bytes())
|
| 85 |
+
report.update(status='candidate_created',candidate_sha256=hashlib.sha256(candidate.read_bytes()).hexdigest(),
|
| 86 |
+
history=[json.loads(line) for line in (out/'log.jsonl').read_text().splitlines()])
|
| 87 |
+
receipt = directory/'receipt.json'
|
| 88 |
+
receipt.write_text(json.dumps(report,indent=2)+'\n')
|
| 89 |
+
return report, str(candidate) if candidate else None, str(receipt)
|