docs: drop coordinator memory below 4GB validator threshold

iris job run rejects --memory >= 4 GB without --enable-extra-resources
(lib/iris/src/iris/cli/job.py:432). The mechanical rewrite produced
--memory=4G, which trips the validator; reduce to --memory=2G to match
the canonical ferries pattern (experiments/ferries/OPS.md:17). Also
correct the stale --memory 16g example in lib/iris/OPS.md:42.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: yonromai

.agents/projects/ferry_framework.md

@@ -274,7 +274,7 @@ gh issue list \
 Launch shape (illustrative, to pin in recipe):
 
 ```bash
-uv run iris --cluster=marin job run --no-wait --cpu=1 --memory=4G --extra=cpu \
+uv run iris --cluster=marin job run --no-wait --cpu=1 --memory=2G --extra=cpu \
   -- python -m experiments.ferries.daily --run_name "daily-125m-$(date +%F)"
 ```
 

.agents/skills/ferries/SKILL.md

@@ -124,7 +124,7 @@ Then push the launch commit (no proposal PR by default).
 Before launch, confirm requester approval in-thread unless they already gave explicit "launch without asking" permission.
 
 ```bash
-uv run iris --cluster=marin job run --no-wait --cpu=1 --memory=4G --extra=cpu \
+uv run iris --cluster=marin job run --no-wait --cpu=1 --memory=2G --extra=cpu \
   -- python -m experiments.ferries.daily
 ```
 
@@ -136,7 +136,7 @@ After launch, capture and post to the issue:
 
 Optional deterministic daily rerun name:
 ```bash
-uv run iris --cluster=marin job run --no-wait --cpu=1 --memory=4G --extra=cpu \
+uv run iris --cluster=marin job run --no-wait --cpu=1 --memory=2G --extra=cpu \
   -e FERRY_DATE "$(date +%Y%m%d-%H%M%S)-daily-ferry" \
   -- python -m experiments.ferries.daily
 ```

docs/explanations/executor.md

@@ -103,7 +103,7 @@ as the entrypoint of a CPU-only Iris job. The script then uses `executor_main`
 to spawn the accelerated sub-jobs via Fray:
 
 ```bash
-uv run iris --cluster=marin job run --cpu=1 --memory=4G --extra=cpu \
+uv run iris --cluster=marin job run --cpu=1 --memory=2G --extra=cpu \
   -- python -m experiments.tutorials.hello_world
 ```
 

docs/recipes/add_scaling_heuristic.md

@@ -42,7 +42,7 @@ Nemotron mix. Then sweep the optimizer hyperparameter space with
 `experiments/references/reference_hyperparameter_sweep.py` or an equivalent setup.
 
 ```sh
-uv run iris --cluster=marin job run --no-wait --cpu=1 --memory=4G --extra=cpu \
+uv run iris --cluster=marin job run --no-wait --cpu=1 --memory=2G --extra=cpu \
   -e WANDB_API_KEY "$WANDB_API_KEY" \
   -- python -m experiments.references.reference_hyperparameter_sweep
 ```
@@ -174,7 +174,7 @@ SCALING_SUITES = {
 Submit:
 
 ```sh
-uv run iris --cluster=marin job run --no-wait --cpu=1 --memory=4G --extra=cpu \
+uv run iris --cluster=marin job run --no-wait --cpu=1 --memory=2G --extra=cpu \
   -e WANDB_API_KEY "$WANDB_API_KEY" \
   -- python -m experiments.isoflop_sweep
 ```

docs/tutorials/train-an-lm.md

@@ -151,7 +151,7 @@ Iris entrypoint job. `executor_main` inside the script spawns the TPU/GPU
 sub-tasks via Fray:
 
 ```bash
-uv run iris --cluster=marin job run --cpu=1 --memory=4G --extra=cpu \
+uv run iris --cluster=marin job run --cpu=1 --memory=2G --extra=cpu \
   -e WANDB_API_KEY "$WANDB_API_KEY" \
   -- python -m experiments.${YOUR_EXPERIMENT_SCRIPT}
 ```

docs/tutorials/train-dpo.md

@@ -141,7 +141,7 @@ Submit the job to the shared Iris cluster (CPU-only entrypoint; the script's
 `executor_main` spawns the TPU sub-task via Fray):
 
 ```bash
-uv run iris --cluster=marin job run --no-wait --cpu=1 --memory=4G --extra=cpu \
+uv run iris --cluster=marin job run --no-wait --cpu=1 --memory=2G --extra=cpu \
   -e WANDB_API_KEY "$WANDB_API_KEY" \
   -- python -m experiments.my_dpo_experiment
 ```

experiments/README_sft.md

@@ -11,7 +11,7 @@ The default doc reproduces OLMO SFT
 Run Olmo SFT with:
 
 ```bash
-uv run iris --cluster=marin job run --cpu=1 --memory=4G --extra=cpu \
+uv run iris --cluster=marin job run --cpu=1 --memory=2G --extra=cpu \
   -e HF_TOKEN "$HF_TOKEN" \
   -- python -m experiments.exp227_sft
 ```
@@ -45,12 +45,12 @@ In `train_step`, essential parameters:
 
 ```bash
 # Basic run
-uv run iris --cluster=marin job run --cpu=1 --memory=4G --extra=cpu \
+uv run iris --cluster=marin job run --cpu=1 --memory=2G --extra=cpu \
   -e HF_TOKEN "$HF_TOKEN" \
   -- python -m experiments.my_sft
 
 # Force specific steps
-uv run iris --cluster=marin job run --cpu=1 --memory=4G --extra=cpu \
+uv run iris --cluster=marin job run --cpu=1 --memory=2G --extra=cpu \
   -e HF_TOKEN "$HF_TOKEN" \
   -- python -m experiments.my_sft --force_run '["your_step_name"]'
 ```

experiments/grug/README.md

@@ -40,7 +40,7 @@ uv run python experiments/grug/base/launch.py
 Iris cluster run (from a dev box, on `marin` prod cluster):
 
 ```bash
-uv run iris --cluster=marin job run --cpu=1 --memory=4G --extra=cpu \
+uv run iris --cluster=marin job run --cpu=1 --memory=2G --extra=cpu \
   -e WANDB_API_KEY "$WANDB_API_KEY" \
   -- python -m experiments.grug.base.launch
 ```

experiments/rollout_data/coderforge.py

@@ -4,7 +4,7 @@
 """togethercomputer/CoderForge-Preview rollout dataset.
 
 Usage:
-    uv run iris --cluster=marin job run --cpu=1 --memory=4G --extra=cpu \
+    uv run iris --cluster=marin job run --cpu=1 --memory=2G --extra=cpu \
       -- python -m experiments.rollout_data.coderforge
 """
 

experiments/rollout_data/gpt_oss_rollouts.py

@@ -4,7 +4,7 @@
 """andyrdt/gpt-oss-20b-rollouts rollout dataset (non-benchmark subsets).
 
 Usage:
-    uv run iris --cluster=marin job run --cpu=1 --memory=4G --extra=cpu \
+    uv run iris --cluster=marin job run --cpu=1 --memory=2G --extra=cpu \
       -- python -m experiments.rollout_data.gpt_oss_rollouts
 """