Skip calibrating with generated tokens in the calibration loop.

[abhinaykukkadapu]

This is to track the optimization after skipping _generate and for context, see the parent issue.

During prompt calibration, after prefilling the prompt tokens into the KV cache, _generate runs an autoregressive loop:

while total_token_list[-1] != tokenizer.eos_id and num_tokens < max_seq_len:
    # generate one token per forward pass

For a 16-token prompt with max_seq_len=1024, this produces 546 forward passes before hitting EOS, all wasted work since quantization observers already have sufficient activation statistics from the prefill pass.

Note: Task calibration is unaffected, wikitext chunks fill the context window (1023 of 1024 tokens), so _generate would exit after at most 1 step.

Both runs: max_seq_len=1024, prefill_ar_len=128, --tasks wikitext --limit 1 + user prompt.

Phase	Baseline (min)	skip_generate=True (min)	Saved
DECODE
calibration (tasks)	122.6	89.1	— (variance)
calibration (prompts) (546 _generate fwd passes)	37.1	0.5	36.6 min
PREFILL
calibration (tasks)	5.8	5.4	—
calibration (prompts) (546 _generate fwd passes)	152.0	0.3	151.7 min
Lowering + QNN Compile
qnn_manager.Compile	113.3	102.4	— (variance)
Total end-to-end	484 min (8.1h)	235 min (3.9h)	249 min (4.2h)

Prompt calibration savings: 188.3 min (3.1h) — from 189.1 min down to 0.8 min.

Projected savings for qwen3-1_7b

Phase	Before (min)	After (min)	Saved
DECODE prompt calibration	40.3	~0.8	39.5
PREFILL prompt calibration	280.5	~0.2	280.3
Total	653 (10.9h)	~333 (5.6h)	~320 (5.3h)

---

cc @cccclai @winskuo-quic @shewu-quic @haowhsu-quic @DannyYuyang-quic @cbilgin