Agent Skills: HuggingFace Transformers Generation Internals

HuggingFace Transformers Generation Internals

Reference for HF Transformers generation pipeline internals, verified against transformers 5.3.0 source code. Covers the full path from model.generate() to output tensors.

Extension Files

| File | Content | |------|---------| | references/generate-loop.md | generate() entry point, mode dispatch, _sample() loop structure, token selection, prefill | | references/scores-and-logits.md | output_scores vs output_logits, tensor shapes, float32 conversion, GenerateDecoderOnlyOutput, memory, compute_transition_scores() | | references/logits-processors.md | LogitsProcessorList, processing order, always-active vs sampling-only processors, key implementations | | references/kv-cache.md | Cache class hierarchy, DynamicCache, attention mask / position_id updates, kvpress interaction | | references/stopping-criteria.md | StoppingCriteria interface, built-in criteria, custom stopping, when checked in the loop | | references/chat-templates.md | apply_chat_template(), Jinja2 rendering, add_generation_prompt, continue_final_message, tokenization patterns |

Quick Decision Tree

What do you need from generation?
│
├─ Per-token logit features    → output_scores=True (see scores-and-logits.md)
│  └─ Need truly raw logits?   → output_logits=True (bypasses LogitsProcessors)
│
├─ Custom stopping logic       → StoppingCriteria subclass (see stopping-criteria.md)
│
├─ Modify token probabilities  → LogitsProcessor subclass (see logits-processors.md)
│
├─ KV cache compression/debug  → past_key_values / DynamicCache (see kv-cache.md)
│
├─ Chat-formatted prompts      → tokenizer.apply_chat_template() (see chat-templates.md)
│
└─ Understand the loop itself  → _sample() internals (see generate-loop.md)

Critical Facts

1. scores != logits: .scores are post-LogitsProcessor, .logits are raw model output. Both are pre-softmax. With do_sample=False and no custom processors, they're identical.

2. Greedy and sampling share _sample(): In v5.3.0, there is no separate _greedy_search(). The do_sample flag controls behavior inside the same method.

3. Always float32: Logits are converted to float32 before processing (line 2754), regardless of model precision (float16/bfloat16).

4. Score shape: Each element of outputs.scores is (batch_size, vocab_size) — one tensor per generated token.

5. Memory cost: For Qwen2.5-7B (vocab=152064), 512 tokens of scores costs ~296 MB per batch element in float32.

6. Sampling processors are gated: Temperature, top-k, top-p warpers are only added when do_sample=True. With greedy decoding, scores pass through the pipeline essentially unmodified.

7. Left-pad required: Decoder-only models require left-padding for batched generation. HF warns if right-padding is detected.

Review Checklist

□ Using return_dict_in_generate=True with output_scores=True?
□ Indexing scores correctly? scores[t] is (batch, vocab), not (vocab,)
□ Know whether scores are pre- or post-processor for your use case?
□ Not holding all score tensors on GPU unnecessarily? (move to CPU or process inline)
□ Padding direction correct for batched generation? (left-pad for decoder-only)
□ StoppingCriteria returns BoolTensor of shape (batch,)?
□ Custom LogitsProcessor signature: (input_ids: LongTensor, scores: FloatTensor) -> FloatTensor?
□ Chat template rendered with add_generation_prompt=True for generation?