diffuse-coordinator job create

Start a LoRA fine-tuning run.

Synopsis

diffuse-coordinator job create [OPTIONS]

Options

flag	value	default	description
--config	<CONFIG>	$DIFFUSE_COORDINATOR_CONFIG	Configuration file. Its [admin] section says where to connect
--endpoint	<ENDPOINT>	$DIFFUSE_COORDINATOR_ENDPOINT	Coordinator endpoint, e.g. https://coordinator.internal:7443
--ca-cert	<CA_CERT>	$DIFFUSE_CA_CERT	The deployment CA certificate (PEM)
--cert	<CERT>	$DIFFUSE_CERT	This process's certificate chain (PEM)
--key	<KEY>	$DIFFUSE_KEY	This process's private key (PEM)
--model	<MODEL>	-	The base model, as shown by model list
--dataset	<DATASET>	-	The training data, as shown by dataset list
--pool	<POOL>	-	Which pool to place it in. Every healthy node by default
--as	<ADAPTER_KEY>	-	What to call the adapter. Derived from the model and the data by default
--rank	<RANK>	16	LoRA rank. Higher learns more and costs more
--alpha	<ALPHA>	32	LoRA alpha; the merged delta is scaled by alpha/rank
--targets	<TARGETS>	q_proj,v_proj	Which projections to adapt, comma separated
--learning-rate	<LEARNING_RATE>	0.0001	Optimiser step size. Lower it if the loss moves erratically rather than settling; raise it only with a suite to check the result against
--batch	<BATCH>	1	Examples per step. The activation term scales with this
--max-seq-len	<MAX_SEQ_LEN>	512	Tokens per example. Longer examples are truncated, which changes what is learned: the run says so when it happens
--epochs	<EPOCHS>	1	Passes over the corpus. More than three on a small corpus usually memorises it: eval against a held-out suite is how you tell
--max-steps	<MAX_STEPS>	0	Stop after this many steps, whatever the data says. Zero means one pass
--seed	<SEED>	0	Seed for shuffling and initialisation. Zero picks one and records it on the job, so a run is reproducible without having chosen to be
--gradient-checkpointing	flag	-	Trade compute for memory: about 40% of the activation term, at roughly 30% more time per step
--checkpoint-every-steps	<CHECKPOINT_EVERY_STEPS>	20	Write a checkpoint every N steps. A machine taken back costs at most this many steps

Notes

The explicit path, when you want to choose the hyperparameters yourself.

Examples

$ diffuse-coordinator job create --model qwen2.5-3b --dataset berichte --as berichte-v1 --rank 32 --epochs 4

job 4f2a9c created and queued.

---

← diffuse-coordinator job · All commands