Add Granite 4.1 20B support (GraniteSWA adapter)

[lipikaworkemail-ctrl]

Summary

Adds support for Granite 4.1 20B (ibm-research/granite-4.1-20b) via a new hf_granite_swa.py adapter for the GraniteSWAConfig model family.

Granite 4.1 introduces a Sliding Window Attention (SWA) architecture where decoder layers alternate between full attention and sliding-window attention (layer_type on each GraniteSWADecoderLayer). Sliding-window layers restrict attention to a local band of sliding_window tokens; full-attention layers use the standard causal mask. This requires a dedicated adapter rather than reusing hf_granite.py.

Changes

New file

• hf_adapters/hf_granite_swa.py — Spyre adapter for GraniteSWAForCausalLM:
  • _make_compiled_block: per-layer compiled forward; builds a band mask for sliding-window layers, passes the normal causal mask for full-attention layers.
  • _run_forward: embedding × multiplier → blocks → RMSNorm → LM head with logits_scaling.
  • prepare_for_spyre: patches RMSNorm, pads LM head, compiles all blocks.

Modified files

• hf_adapters/auto_spyre_model.py
  • Import GraniteSWAConfig and hf_granite_swa.
  • Register GraniteSWAConfig → hf_granite_swa in CONFIG_TO_ADAPTER_MODULE_MAPPING.
  • Register "ibm-research/granite-4.1-20b": torch.bfloat16 inMODEL_PATH_TO_TORCH_DTYPE.
• tests/model_registry.py
  • Add granite41_20b (ibm-research/granite-4.1-20b, hf_granite_swa.py, 20b) — verified.
• ARCHITECTURE.md — verified checkpoint row, adapter table row, coverage counts, model-specific features column for Granite 4.1 SWA.
• hf_adapters.md — adapter summary row .
• .github/cache_config/hf_models_and_datasets.yaml — add ibm-research/granite-4.1-20b.

Verified checkpoint

Model	Repo	dtype	Prompt	Output
Granite 4.1 20B	ibm-research/granite-4.1-20b	bfloat16	"The capital of France is"	Paris.

Tested on Spyre with max_new_tokens=8.

Notes

• ibm-research/granite-4.1-20b lives under the ibm-research org (notibm-granite). All other Granite 4.1/4.0/3.x models remain under ibm-granite.
• The aten.embedding CPU fallback warning on Spyre is expected and does not affect correctness.
• int64→int32 downcast warning is backend-level and expected.

Checklist

• New adapter follows prepare_for_spyre / _run_forward / _make_compiled_block pattern
• GraniteSWAConfig registered in CONFIG_TO_ADAPTER_MODULE_MAPPING
• MODEL_PATH_TO_TORCH_DTYPE updated with correct org and dtype for 20B
• Only verified checkpoint (20B) added to ARCHITECTURE.md and hf_adapters.md
• Coverage counts updated (28 adapters · 45 verified checkpoints · 28 generative)
• Model-specific features table updated in ARCHITECTURE.md
• Cache config updated with ibm-research/granite-4.1-20b

Results:

[lipikaworkemail-ctrl]

Hi @arielge, could you please review this PR?

cc: @saurabh-srivastava4 @goutham-binnadi-gopala @Mohini-Shrivastava

[arielge]

Thanks @lipikaworkemail-ctrl, I think that to merge this we would need to wait for a transformers release that includes the new GraniteSWAConfig class, and upgrade the dependencies accordingly

[arielge]

Suggested change

	"adapter": "hf_granite_swa.py",
	"adapter": "hf_granite_swa.py",
	"is_gated": True,

until it is public, let's consider it gated

[arielge]

Maybe better to do more reuse of the existing adapter, like in #281

[arielge]

The import currently fails, so to merge this change we would need to wait for a transformers release that includes this config