Add a SmolLM2 compatibility smoke result and support matrix

[DaBestCode]

Prove the existing Transformers adapter against a second public full-attention decoder family and turn compatibility into explicit documentation.

Suggested checkpoint: HuggingFaceTB/SmolLM2-135M (a different checkpoint is acceptable after discussion).

Acceptance criteria:

• Run scripts/smoke_transformers.py with enough generated tokens to trigger at least two eviction rounds.
• Check in the raw JSON result with exact model revision, environment, seed, budget, and buffer size.
• Add a README compatibility table that distinguishes verified, expected, and unsupported model families.
• Do not add model weights or run checkpoint downloads in CI.
• Do not make a throughput or quality claim from the smoke result.
• Run the complete local verification suite.

Comment with the proposed checkpoint before starting so we do not duplicate model coverage.

[RezaAshrafii]

Hi, I'd like to work on this issue.

I can use HuggingFaceTB/SmolLM2-135M as the proposed checkpoint.

My plan is to:

1. Run the existing scripts/smoke_transformers.py with enough generated tokens to trigger at least two eviction rounds.
2. Record the raw JSON result with the exact model revision, environment, seed, budget, and buffer size.
3. Add a README compatibility table distinguishing verified, expected, and unsupported model families.
4. Keep model downloads out of CI and avoid making throughput or quality claims from the smoke result.
5. Run the complete local verification suite before opening a pull request.

Please let me know if this scope looks good and whether this checkpoint would duplicate any planned coverage.

[DaBestCode]

Thanks for volunteering — assigned to you.

HuggingFaceTB/SmolLM2-135M is a good choice here. I checked its config: it is a standard Llama causal LM with full attention, GQA (9 query heads / 3 KV heads), and no sliding-window setting, so it should exercise the adapter without hitting the currently unsupported hybrid-attention path.

Please record the exact model revision and environment in the fixture. For the smoke run, use a cache budget of 64, recency buffer of 8, and enough generated tokens to observe at least two eviction rounds. The acceptance criteria you summarized are exactly right: committed raw JSON plus the compatibility-table row, no downloaded weights or performance claim, and the full test suite passing.

Feel free to open a draft PR early if you want feedback on the fixture shape before completing the run.