Spaces:

Quazim0t0
/

DaisyChain-Web

Sleeping

App Files Files Community

DaisyChain-Web

680 kB

Ctrl+K

1 contributor

History: 21 commits

Quazim0t0

FineWeb-Edu parquet streaming: /data endpoint (hyparquet over CDN range requests)

d446697 verified about 2 months ago

public
Dispatch-optimized backward (bit-identical) + IEEE-754 oracle + oracle mutation tests + TEST_RESULTS.md about 2 months ago
.gitattributes

1.52 kB
initial commit about 2 months ago
.gitignore

30 Bytes
DaisyChain-Web demo: P2P browser training through verified INT8 units about 2 months ago
Dockerfile

182 Bytes
Room-first lobby (Create/Join, host approval), group training settings sliders about 2 months ago
LICENSE

1.09 kB
DaisyChain-Web demo: P2P browser training through verified INT8 units about 2 months ago
README.md

4.84 kB
README: front matter + full body (front matter must never be dropped) about 2 months ago
TEST_RESULTS.md

5.26 kB
Dispatch-optimized backward (bit-identical) + IEEE-754 oracle + oracle mutation tests + TEST_RESULTS.md about 2 months ago
package-lock.json

867 Bytes
DaisyChain-Web demo: P2P browser training through verified INT8 units about 2 months ago
package.json

579 Bytes
Dispatch-optimized backward (bit-identical) + IEEE-754 oracle + oracle mutation tests + TEST_RESULTS.md about 2 months ago
server.js

8.84 kB
FineWeb-Edu parquet streaming: /data endpoint (hyparquet over CDN range requests) about 2 months ago
test_core.js

2.06 kB
DaisyChain-Web demo: P2P browser training through verified INT8 units about 2 months ago
test_corpus.js

10.9 kB
Dispatch-optimized backward (bit-identical) + IEEE-754 oracle + oracle mutation tests + TEST_RESULTS.md about 2 months ago
test_gates.js

5.54 kB
Exact kernel gates: verify variants emitted from the same WGSL source restore the int32 oracle on the kernels training actually runs; bit-exact epilogue mirror removes the tolerance entirely; shape sweep; dead exactly-gated kernel deleted. Continuous verification: random-cell audit at live shapes + canonical cross-device kernel probe (the weight hash structurally cannot detect kernel disagreement). New test_gates.js mutation-tests the gate itself. about 2 months ago
test_ieee.js

11.4 kB
Dispatch-optimized backward (bit-identical) + IEEE-754 oracle + oracle mutation tests + TEST_RESULTS.md about 2 months ago
test_metamorphic.js

9.74 kB
Dispatch-optimized backward (bit-identical) + IEEE-754 oracle + oracle mutation tests + TEST_RESULTS.md about 2 months ago
test_optimizer.js

3.42 kB
DaisyChain-Web demo: P2P browser training through verified INT8 units about 2 months ago
test_transformer.js

2.19 kB
Web trainer upgrade: sync guard (leader roster + weight-hash divergence check), 3xINT8 fast-accurate GEMM, DP4A hardware path (verified vs units), Spikewhale tokenizer (16.5k vocab), FineWeb-Edu streaming, gradient/checkpoint fragmentation, inference kit, generation tester, contribution logs about 2 months ago
test_unit_backward.js

3.62 kB
Default learning rate 0.02 -> 0.01 (200-step sweep on the real 16512 vocab: 0.02 gives 1.93 at 100 steps vs 1.01 for 0.01; they converge by 200 but short runs are what the demo is). Batch the per-head backward GEMMs: 4 batched dispatches per layer instead of 4 per head, GEMM calls 167 -> 47 per step, float backward ~14% faster. Adds test_unit_backward.js, the experiment showing int8 gradients do not damage convergence. about 2 months ago
test_verified.js

2.96 kB
DaisyChain-Web demo: P2P browser training through verified INT8 units about 2 months ago