[quantization] Support multi-turn workflows

[mhs4670go]

This commit supports multi-turn workflows.

python -m tico.quantization.recipes.debug.static_llama_runtime \
  --model /mnt/sdb/model/Llama-3.2-1B-Instruct \
  --max-seq 256 \
  --device cpu \
  --no-run-baseline \
  --run-multiturn \
  --append-bucket 16


python -m tico.quantization.recipes.debug.static_llama_runtime \
  --model /mnt/sdb/model/Llama-3.2-1B-Instruct \
  --max-seq 256 \
  --padding-side right \
  --device cpu \
  --no-run-baseline \
  --run-multiturn \
  --multiturn-first-prompt $'User: What is the capital of France?\nAssistant:' \
  --multiturn-second-turn $'\nUser: What country is that city in?\nAssistant:' \
  --multiturn-decode-steps-before-append 4 \
  --multiturn-decode-steps-after-append 4 \
  --append-bucket 16 \
  --compare-append-prefill-with-decode-loop

...
====================================================================================================
Multi-turn post-append decode logits step 1
mean|diff| = 0.02939297
 max|diff| = 0.10098839
PEIR       = 0.340664 %
====================================================================================================
Multi-turn post-append decode logits step 2
mean|diff| = 0.03495336
 max|diff| = 0.17215163
PEIR       = 0.569002 %
====================================================================================================
Multi-turn post-append decode logits step 3
mean|diff| = 0.02905156
 max|diff| = 0.19826031
PEIR       = 0.583857 %
====================================================================================================
Multi-turn post-append decode logits step 4
mean|diff| = 0.01603191
 max|diff| = 0.09334183
PEIR       = 0.341133 %
====================================================================================================
Multi-turn generated text:
User: What is the capital of France?
Assistant: The capital of France
User: What country is that city in?
Assistant: France
User:

Related: #812
TICO-DCO-1.0-Signed-off-by: seongwoo mhs4670go@naver.com

[dvsav]

Couldn't find any references to this class.

[dvsav]

This class looks absolutely identical to LlamaDecoderLayerDecodeExportAdapter. Can we simply reuse the latter one?

[mhs4670go]

Agreed. I'll remove the duplicate class and reuse it while keeping append_prefill as a separate export mode.

[dvsav]

The actual code returns a flat tuple (hidden_states, new_key, new_value), not a nested tuple.

[dvsav]

(Optional) Add unit tests

The PR adds a lot of new functionality (new export mode, new adapters, multi-turn runtime with attention mask construction, cache management, chunked append) but includes no unit tests. The only validation is the debug script static_llama_runtime.py. Key areas needing tests:

• _build_append_prefill_attention_mask correctness (causal within new block, masking of gap, padded rows).
• append_prefill cache write-back correctness.
• append_prefill_chunked with chunks smaller than bucket size.
• Export of append-prefill graphs at various bucket sizes.

[dvsav]

(Optional) return_type = "tuple" mutation in __init__ is fragile

All three decoder-layer adapters set self.wrapped.return_type = "tuple" in __init__. Since the runtime creates all three adapters from the same wrapped module, the last __init__ wins. This works today because all set the same value, but it's a latent bug if a future adapter sets a different return_type. Consider setting return_type in forward() or making it a parameter of the adapter rather than mutating the wrapped module.

[dvsav]

(Optional) Implement bucket selection logic

The issue specification describes runtime bucket selection: "The runtime selects the smallest usable bucket satisfying: T <= Q and P + Q <= K." The current append_prefill_chunked uses a fixed bucket_size for all chunks. A production runtime should select the smallest bucket that fits each chunk, and handle the case where no bucket fits (context-length policy). The current implementation (_normalize_append_prefill_buckets) raises ValueError when capacity is exceeded, which is acceptable for a first version but should be documented as a limitation.

[mhs4670go]

@dvsav PTAL:)

[dvsav]

👍 LGTM. Thank you!

[Torrero]

LGTM

[mhs4670go]

@dvsav @Torrero I've just rebased the PR. PTAL

[dvsav]

👍