IndependentMonarch/vllm-state-transfer

PVTS: probe transfer from HuggingFace to vLLM

Does a bi-encoder probe trained on hidden states harvested with HF transformers still work when the states come from a vLLM engine? Each cell scores the same trained probe twice over the same test rows, varying only the source of the query vector. Tool keys are HF-derived and held fixed for both arms. No retraining, no threshold refitting.

Within noise in every cell: the probe transfers to vLLM unmodified.

model	dataset	layer	n	cosine	norm_ratio	acc_hf	acc_vllm	delta	se	agreement	margin_ratio
llama-3.2-3b	xlam-human-gold	16	3447	0.9990	1.0060	0.9626	0.9626	0.0000	0.0032	0.9977	1.0001
llama-3.2-3b	xlam-self-distill	16	3879	0.9989	1.0065	0.8580	0.8639	0.0059	0.0056	0.9881	1.0003
mistral-7b-v0.3	xlam-human-gold	16	3447	0.9999	1.0003	0.9539	0.9542	0.0003	0.0036	0.9991	1.0000
mistral-7b-v0.3	xlam-self-distill	16	5756	0.9999	1.0003	0.7835	0.7830	-0.0005	0.0054	0.9953	0.9995

cosine and norm_ratio are wiring diagnostics over the first 64 rows: norm_ratio near 1 confirms the final-norm patch took, cosine near 1 confirms the read position and tokenization match harvest. They fail independently. margin_ratio is the median per-example ratio of top-2 logit margins — near 1 means the selective-prediction thresholds fit on HF margins carry over to vLLM without refitting.

Files

• summary.csv — one row per (model, regime)
• vllm_transfer_{model_id}__{dataset_id}.csv — per example: pred_hf, pred_vllm, conf_*, margin_*, gold, n_cand, n_tokens
• manifest.json — source repos, config, and stack versions

Method notes

The vLLM stack is truncated to the probe's best_layer by writing a physically truncated checkpoint (hf_overrides fails: vLLM's loader raises on checkpoint weights with no module to receive them). model.norm is retained in the checkpoint and neutralized at runtime: Llama-family decoder layers return (hidden_states, residual) and model.norm is where the add happens, so it is replaced with an add-without-norm module rather than an identity. Pooling is LAST with use_activation=False; prompts are cut to ids[:eou+1] and passed as token ids, so the pooled vector is the same one harvest read at eou.

Prompts are tokenized with add_special_tokens=True, matching harvest. This is load-bearing for Mistral: the [/INST] position finder computes its index against a no-special-tokens tokenization, so under harvest's convention it returns a position one token early. Train and eval are consistent, so the published Mistral accuracies stand, but the read is on the [/INST] interior rather than its closing token. Reproducing the offset is required; "fixing" it here would put train and eval on different tokens and costs ~8 accuracy points.

Stack: vLLM 0.27.1, torch 2.13.0+cu130 (CUDA 13.0), NVIDIA RTX PRO 6000 Blackwell Server Edition. These figures are specific to that stack.