Address review on 1429: skill frontmatter, TP-formula consistency, docs

- eagle3-new-model: add user_invocable:true (match sibling eagle3 skills);
  fix internally-inconsistent GPU-sizing table (tp is fixed at 4, full-node
  sharding; gpus_to_fit only sizes node count) and use consistent example rows;
  update vLLM backend note to native extractor (no speculators).
- eagle3-review-logs: correct "in parallel" wording (the loop is sequential).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Ye Yu <yeyu@nvidia.com>

Author: yeyu-nvidia

.claude/skills/eagle3-new-model/SKILL.md

@@ -6,6 +6,7 @@ description: >
   backend (TRT-LLM / HF / vLLM) and GPU configuration.
   Use when user wants to run EAGLE3 on a model that does not yet have a YAML in
   tools/launcher/examples/ or asks how to configure the pipeline for a new checkpoint.
+user_invocable: true
 ---
 
 # EAGLE3 New Model Configuration
@@ -31,24 +32,27 @@ Determine these values from the HuggingFace model card, `config.json`, and vLLM
 OCI-HSG nodes: **4 GPUs × 192 GB HBM3e = 768 GB per node**
 
 ```text
-BF16 weight size  = total_params × 2 bytes
-GPUs needed       = ceil(weight_size_GB / 192)
-nodes             = ceil(gpus_needed / 4)
-tp                = min(gpus_needed, 4)
+weight_size_GB = total_params × 2 bytes        # BF16
+gpus_to_fit    = ceil(weight_size_GB / 192)     # min GPUs to hold weights (192 GB each)
+nodes          = ceil(gpus_to_fit / 4)          # 4 GPUs per node
+tp             = 4                              # shard across all 4 GPUs on each node
 ```
 
-| Model | Weights (BF16) | GPUs | nodes | tp |
+`tp` is fixed at 4: jobs allocate whole nodes and shard across the node's 4 GPUs, with
+data parallelism across nodes (SLURM array tasks). `gpus_to_fit` is only used to size `nodes`.
+
+| Model | Weights (BF16) | GPUs to fit | nodes | tp |
 |---|---|---|---|---|
 | 8B dense | ~16 GB | 1 | 1 | 4 |
 | 70B dense | ~140 GB | 1 | 1 | 4 |
-| 685B MoE | ~340 GB | 2 | 1 | 4 |
-| 1T MoE | ~595 GB | 4 | 1 | 4 |
+| 405B dense | ~810 GB | 5 | 2 | 4 |
+| 671B MoE | ~1.3 TB | 7 | 2 | 4 |
 
 ## Step 3 — Choose the hidden state dump backend
 
 | Backend | Script | When to use |
 |---------|--------|-------------|
-| vLLM | `common/eagle3/dump_offline_data_vllm.sh` | Default; broad coverage via vLLM + speculators |
+| vLLM | `common/eagle3/dump_offline_data_vllm.sh` | Default; broad coverage via vLLM's native hidden-state extractor |
 | HF | `common/eagle3/dump_offline_data_hf.sh` | VLMs, custom-code models, SWA attention |
 | TRT-LLM | `common/eagle3/dump_offline_data.sh` | Pure-text models with TRT-LLM support (needs `--tp`/`--moe-ep`) |
 

.claude/skills/eagle3-review-logs/SKILL.md

@@ -31,7 +31,7 @@ Do this in a single Bash call. If no experiments exist, ask the user for the dir
 
 ## Step 1 — Read all task logs
 
-Read the last 200 lines of each log in parallel. Errors appear at the end:
+Read the last 200 lines of each log in a single Bash call. Errors appear at the end:
 
 ```bash
 for f in $(find experiments/<exp_id>/ -name "sbatch_*.out" | sort); do