Gahow Wang

gahow pushed to main at gahow/xtrain

2026-06-16 09:56:32 +00:00

579365f4a0 docs: run v5 — TinyStories saturation at dim768 (val 1.11)

8a1e29543b run: v5 archive + export (dim768, bf16, 5.33ep, val 1.11)

Compare 2 commits »

gahow pushed to main at gahow/xtrain

2026-06-16 06:29:26 +00:00

5b7dde1736 test: bf16 test reads f32-cast logits (forward now returns bf16)

gahow pushed to main at gahow/xtrain

2026-06-16 06:28:27 +00:00

320c1ae4fb perf: KI-2 FIXED — dim768 bf16 fits batch 32, tok/s 31.5K→40.8K

gahow pushed to main at gahow/xtrain

2026-06-16 06:20:50 +00:00

48922cb628 perf: keep bf16 logits (no persistent fp32 logits buffer)

gahow pushed to main at gahow/xtrain

2026-06-16 06:15:09 +00:00

30db62d8f2 docs: Phase T12 — bf16 mixed precision design

0a2a4dcaa8 train: --bf16 flag (fp32-master AMP) + bf16 correctness test

b0086b5214 autodiff: bf16 mixed-precision path (fp32 master via cast op)

d05115ddf3 cuda: bf16 cuBLAS GemmEx (16BF in/out, fp32 accum) + cast kernels

Compare 4 commits »

gahow pushed to main at gahow/xtrain

2026-06-16 05:14:52 +00:00

511ceebbb3 docs: KI-2 trigger — dim768 fp32 batch-32 OOM

ff79fee3c5 docs: run v4 — TinyStories, dim768, val 1.17

734e119db3 run: v4 archive + export (dim768, 8-GPU DDP, val 1.17)

Compare 3 commits »

gahow pushed to main at gahow/xtrain

2026-06-16 03:15:13 +00:00

f85bd4d276 perf: KI-5 FIXED — single-GPU 40K->93K tok/s, DDP scaling 1.3x->5x@8

gahow pushed to main at gahow/xtrain

2026-06-16 03:04:17 +00:00

4c3f332f64 docs: Phase T11 — caching allocator

b7104e2cb7 test: loosen flaky DDP cross-rank assertion to <1e-6; scale to world=8

28801fbfe5 cuda: device caching allocator (pool GpuBuffer alloc)

Compare 3 commits »

gahow pushed to main at gahow/xtrain

2026-06-16 01:42:23 +00:00

d422c68704 docs: KI-5 — correct cross-rank divergence attribution (pre-existing flaky)

gahow pushed to main at gahow/xtrain

2026-06-16 01:40:51 +00:00

84092fb28d docs: KI-5 re-diagnosis — all-reduce is NOT the DDP bottleneck (T11)

88c2c15768 Revert "dist: coalesce grads into buckets for all-reduce (KI-5)"

Compare 2 commits »

gahow pushed to main at gahow/xtrain

2026-06-16 01:09:54 +00:00

b8b58212dc dist: coalesce grads into buckets for all-reduce (KI-5)

gahow pushed to feat/two-stop at gahow/aituner

2026-06-16 01:08:48 +00:00

4a64196a99 Add 27B Stop-B agentic-loop config (harness-driven, GPUs 2-7)

gahow pushed to feat/two-stop at gahow/aituner

2026-06-16 01:08:11 +00:00

b17b213575 Tear down the engine on SIGTERM instead of orphaning it

gahow pushed to main at gahow/xtrain

2026-06-15 19:43:26 +00:00

a78502e0f0 docs: run v3 — TinyStories, dim512, val 1.30

64b2a8c09e run: v3 archive + export (dim512, single-GPU batched, val 1.30)

Compare 2 commits »

gahow pushed to feat/two-stop at gahow/aituner

2026-06-15 17:54:42 +00:00

93ce339d61 Document 27B TP sweep: per-GPU rises sharply with TP (dense), opposite of MoE

gahow deleted branch t10-validate from gahow/xtrain

2026-06-15 16:45:58 +00:00

gahow pushed to main at gahow/xtrain

2026-06-15 16:44:58 +00:00

9a25616a30 docs: Phase T10 — batched forward

4ccab0fb42 perf: KI-1 fixed — GPU util 0-15%→37-54%, tok/s 1653→25627 (15.5x)

25b032445d train: real batched step (drop loop+SUM)

5353b38402 model: batched forward [B,S]

7821bd9c34 autograd: batch dim for ops (flatten linears, batched attention)

Compare 5 commits »

gahow pushed to t10-validate at gahow/xtrain

2026-06-15 16:31:25 +00:00

67687ec8fe wip: T10 batched forward (validation)

gahow created branch t10-validate in gahow/xtrain

2026-06-15 16:19:32 +00:00

gahow pushed to t10-validate at gahow/xtrain

2026-06-15 16:19:32 +00:00

ce9d22ffc2 wip: T10 batched forward (validation)