Fix module-config tensor init: shape-bounded randint high, int-dtype randint, float-dtype xavier, bfloat16 default, int64 for index tensors

[kiszk]

Summary

Fixes four defects in the auto-generated module test configs and the generator/API that produces them (utils/module_discovery/auto_generate_module_config.py). The source fixes make future regenerations correct; the four existing YAMLs under tests/configs/module_tests/ are updated to match.
Fixes #225

Problems & fixes

1. randint upper bound was a fixed 10000

_tensor_info_to_spec set the exclusive upper bound of integer tensors to a fixed 10000, unrelated to the tensor's own shape. For index/position/cache tensors this produces out-of-range values and unexpected element accesses.

Fix (source): high = max(min(shape), 1) — the tensor's smallest dimension, guarded to be ≥ 1.

2. Integer tensors initialized with randn

Some torch.int32/torch.int64 tensors used init: xavier. torch.randn (normal_kernel_cpu) or xavier is float-only and raises NotImplementedError for integer dtypes.

Fix (source): any integer dtype (int/uint/long/short/bool) now always uses randint.

3. Floating-point tensors defaulted to float32 instead of bfloat16

from_pretrained defaults to float32, so models loaded without an explicit dtype recorded float tensors as float32. Spyre runs in bfloat16, so bfloat16 should be the default in both the generator and the API.

Fix (source):

• New DEFAULT_FLOAT_DTYPE = torch.bfloat16 constant.
• API — load_model_only now setdefaults torch_dtype to bfloat16 (explicit caller values still honored).
• YAML generator — _tensor_info_to_spec normalizes any captured float dtype (float16/32/64/half/double) to bfloat16 on emit.

4. Special (index) tensors kept a floating-point dtype

Tensors matched by _is_special_tensor() (name contains position/mask/ids) carry indices, not activations, and are always initialized with randint. Some - notably position_embeddings were captured under a floating-point dtype, leaving the inconsistent combination of a float dtype with randint init.

Fix (source): _is_special_tensor() for position_embeddings returns False.

YAML changes

Applied to all four configs (granite_3_3_8b_instruct, granite_4_1_8b, Ministral-3-14B-Instruct-2512, Mistral-Small-3.2-24B-Instruct-2506):

• Every high: 10000 → the tensor's smallest dimension (cache_position [128] → 128, all others → 1).
• Integer tensors with init: randn → init: randint (+ init_args.high) — hit the RotaryEmbedding inputs in granite_4_1_8b and Mistral-Small.
• Float tensors with init: randn → init: xavier for numerical stability
• Tensor dtype: torch.float32 → torch.bfloat16 — in granite_4_1_8b and Mistral-Small.
• position_embeddings tensor-list entries (float dtype + randint) → init: xavier — 8 per file across all four configs.

Comments, YAML anchors (&id001/*id001), and the supported_dtypes global block are preserved.

[kiszk]

@moriohara @anubhavjana Could you please review this PR?

[moriohara]

lgtm

[spyre-ci]

🔄 merge-queue-integration: running

:hammer_and_wrench: orch *trigger-pr-validation* · arches amd64 · fp amd64=d5ec218f
L0 torch-spyre/torch-spyre-dev amd64 🔨
L1 hf-adapters/hf-adapters-dev amd64 ⏳
L2 spyre-inference/spyre-inference-dev amd64 ⏳