Summary

apr rosetta convert fails when converting GGUF to SafeTensors format due to NaN values detected in tensor data during conversion.

Reproduction

apr rosetta convert \
  ~/.cache/pacha/models/e910cab26ae116eb.gguf \
  /tmp/test.safetensors

Error Output

=== Rosetta Stone Conversion ===

Source: /home/noah/.cache/apr-models/qwen2-5-coder-0-5b-instruct/gguf/model.gguf
Target: /home/noah/.cache/apr-models/qwen2-5-coder-0-5b-instruct/safetensors/model.safetensors

--- Source Inspection ---
  Format: GGUF
  File Size: 491400064 bytes
  Tensors: 291
  Parameters: 630167424
  Architecture: qwen2
  Quantization: 0

Converting...
[GH-185] Embedding 151387 BPE merge rules into APR metadata
error: Validation failed: Conversion failed: Invalid model format: PMAT-187: 
Tensor 'blk.0.attn_k.weight' contains 108 NaN values (data corruption detected). 
Toyota Way: Stop the line - do not pass defects downstream.

Analysis

The NaN values appear during conversion, suggesting one of:

1. Dequantization bug - Q4_K_M or other quantized format produces NaN during float conversion
2. Source corruption - Original GGUF has bad data (unlikely - GGUF→APR works)
3. SafeTensors path bug - Something specific to the SafeTensors writer path

Note: GGUF→APR conversion succeeds for the same source file, which suggests the bug is specific to the SafeTensors conversion path.

Impact

• Blocks 6-column throughput profiling in apr-model-qa-playbook
• SafeTensors format testing impossible
• See: BLOCKED: 6-column throughput profiling requires apr rosetta convert fix apr-model-qa-playbook#2

Environment

apr 0.2.12
Model: Qwen/Qwen2.5-Coder-0.5B-Instruct (from pacha cache)
Source: GGUF (491MB, 291 tensors, qwen2 arch)

Expected Behavior

If source GGUF is valid (which it appears to be since APR conversion works), SafeTensors conversion should also succeed.