weilan55/k3f_gb300_arm_logs

K3-Flash DSpark draft training runs — GB300 cluster logs and configs

Training logs and the config each arm actually trained with for 17 draft-model arms (g3c..g3s) run on a GB300 cluster in 2026-09/10. Archived because the cluster was reclaimed. No checkpoints here — those were ~1.86 TB.

Target model: moonshotai/Kimi-K3-Flash. Trainer: TorchSpec (DSpark / EAGLE-3 family). Corpora: the regenerated datasets published at weilan55/k3f_regen_data, plus internal ones that contained service traffic and were not published.

Arms

loss is the last value the progress line reported; it is a TRAINING loss and is not comparable across corpora of different size or composition.

arm	latest iter	steps	final loss	corpus
g3c	6413	6412	1.771	v3ctl_allturn (pretokenized, all turns supervised)
g3d	651	4580	3.020	v3_train
g3e	4581	4580	1.849	v3_train
g3f	4581	4580	1.507	v3_train
g3g	4581	4580	1.507	v3_train
g3h	2751	4580	1.628	v3_train
g3i	—	4580	—	v3_train (no checkpoint written)
g3j	4581	4580	1.309	v3_train
g3k	2947	—	1.989	v4_train
g3l	7527	—	1.658	v34_train (v3 ∪ v4)
g3m	4455	—	1.462	v3_train
g3n	4455	—	1.621	v3_train
g3o	4455	—	1.632	v3_train
g3p	4581	—	1.417	v5_train
g3q	9161	—	1.277	v5_train
g3r	14325	—	1.261	v6_text_train
g3s	18019	—	1.236	v7_text_train

INDEX.json carries the same table machine-readably, with each arm's max_seq_length, num_epochs, dflash_block_size and last_turn_loss_only.

Layout

k3flash-dspark-<arm>/
  config.yaml                          the config the run actually used
  train.log                            driver log (progress line is \r-delimited)
  attempt.log                          launcher attempts
  latest_checkpointed_iteration.txt    terminal iteration

A few arms carry extra logs named for what happened to them: g3g/attempt_hang_11-24.log, g3g/train_3attempts_node34.log, g3h/train_paused_at_2841.log.

Reading these logs

Training loss does not rank arms across corpora. g3s has the lowest loss (1.236) and the largest corpus (v7, 1.22M rows); g3d has the highest (3.020) and stopped at step 651. Neither fact is an acceptance-rate result. Draft quality here is measured by acceptance length (tau) on held-out serving workloads, which these logs do not contain.

Several arms stopped short of their horizon, by node failure rather than by design: g3c was OOM-killed at step 4893 when a side job filled host RAM, then resumed and completed at 6413; g3h stopped at 2751 of 4580; g3d at 651. The attempt.log files show the restarts.

One node had a faulty GPU. Xid 109 CTX SWITCH TIMEOUT on node 10.0.3.34, reproducible on every vLLM launch. In the application log it surfaces only as No available shared memory broadcast block found in 60 seconds — an INFO line that also prints during a legitimate cold compile. The distinguishing signal is per-GPU utilisation (one rank at 0% while the rest pin at 100%) plus dmesg -T | grep Xid, not the application log.

Corpus caveat for anyone comparing g3k/g3l against a v3_test number. v4_train overlaps v3_test on 149,114 of its 188,599 rows (79%), discovered after these runs. An arm trained on v4 and evaluated on v3_test is reading its own training data.