GPT-2 hidden_dim metadata reports 64 (head_dim) instead of 768

[noahgift]

Bug Report

Source: tiny-model-ground-truth parity checker (0/59 passing)
Severity: Critical — blocks ALL GPT-2 inference
Related: Follow-up to #233 (GPT-2 architecture support)

Description

After the GH-233 fix added GPT-2 architecture detection and tensor name mapping, inference still fails because the metadata reports hidden_dim=64 instead of hidden_dim=768. This is GPT-2's head_dim (768 / 12 heads = 64), not the actual hidden dimension.

The contract validator then expects embeddings of shape [50257, 64] (3.2M elements) but finds [50257, 768] (38.6M elements).

Error Output

[APR-LOAD] Embedding tensor 'model.embed_tokens.weight': dims=[50257, 768], expected [vocab=50257, hidden=64]
[APR-LOAD] Token 0 embedding sample: [-0.1101, -0.0393, 0.0331, 0.1338, -0.0485]  ← good data
[APR-LOAD] Embedding loaded: 38597376 elements (vocab=50257 x hidden=64)  ← 38M is correct for 50257×768

error: F-LAYOUT-CONTRACT-001 Tensor 'token_embedding': Shape mismatch: got 38597376 elements, expected 3216448 (50257x64)

Root Cause

The infer_architecture() or metadata extraction path for GPT-2 is reading n_embd_head_k or similar per-head dimension field instead of n_embd (768). GPT-2 config:

Field	Value
n_embd (hidden_dim)	768
n_head	12
n_embd / n_head (head_dim)	64

The import pipeline is storing 64 as hidden_dim in APR metadata.

Affected Models

Model	Expected hidden_dim	Got hidden_dim	Architecture
GPT-2 124M	768	64	GPT-2

Both Int4 and Int8 are affected.

Fix

In the GPT-2 metadata extraction path (likely infer_architecture() or wherever hidden_dim is read from SafeTensors/HF config), ensure n_embd (768) is used, not n_embd / n_head (64).

Reproduction

cd tiny-model-ground-truth
make clean && make convert
apr run models/gpt2-124m-int4.apr -p "Hello" -n 32 --json
# → expected [vocab=50257, hidden=64] but tensor has 768 columns

Environment

• apr v0.2.16 + Int8 quantization corrupts embedding tensors (NaN/Inf + shape mismatch) #231/232/233 fixes applied
• Platform: Linux x86_64

[noahgift]

Fixed by GH-234 (commit e0e6be6). load_model_config_from_json() now parses GPT-2's n_embd field (falls back from hidden_size), n_head (falls back from num_attention_heads), and n_inner (falls back from intermediate_size, defaults to 4×n_embd). Verified by test_gh235_load_config_gpt2_n_embd_field and test_gh235_load_config_gpt2_n_inner_fallback regression tests.