File size: 5,774 Bytes
15a4381
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
# Checkpoint-to-Model Diagnostic Matching Matrix

## 1. Executive Summary

This report delivers a forensic audit and diagnostic loading matrix of the checkpoint file `qtensorformer_small.pt` against all candidate model architectures in the repository. Under the scientific mandate `strict=True`, zero candidates achieve checkpoint compatibility.

---

## 2. Status Distinctions

### PROVEN FACT
1. `qtensorformer_small.pt` is a serialized PyTorch checkpoint dictionary containing 6 top-level keys:
   - `model_state_dict` (113 tensor weights, total 1,347,747 parameters)
   - `config` (`{'vocab_size': 10000, 'hidden_dim': 128, 'intermediate_size': 256, 'n_heads': 4, 'n_layers': 3, 'seq_len': 64, 'tt_rank': 4, 'n_qubits': 4, 'quantum_layers': 2, 'quantum_ratio': 0.3, 'use_quantum_attention': True, 'use_adaptive_rank': True, 'use_selective_routing': True, 'r_min': 2, 'r_max': 12, 'alpha': 1.0}`)
   - `params`: `1347747`
   - `baseline_params`: `16146432`
   - `compression_ratio`: `11.980313812607262`
   - `version`: `"1.0.0"`
2. The state dictionary relies on TensorLy tensor decomposition layers:
   - `blocks.i.ffn.layer.weight.factors.factor_0..5`
   - `blocks.i.ffn2.layer.weight.factors.factor_0..5`
   - `lm_head.layer.weight.factors.factor_0..5`
   - `embeddings.quantum_enhancer.project_up.weight`
   - `blocks.i.attention.quantum_encoder.project_up.weight`
   - `blocks.i.attention.alpha_gate.0.weight`
3. Across all candidate model implementations in the repository:
   - Common keys with `qtensorformer_small.pt`: **0**
   - Unexpected keys from `qtensorformer_small.pt`: **113** (100% of checkpoint keys)
   - Strict load result: **FAIL** on 100% of candidate models.
4. Git commit history proves `qtensorformer_small.pt` was committed in commit `82fd973cb44f62a21ac0fc0fdfa1b30a8c598754` on May 6, 2026 by `ml-intern`, originating from an external `/app/qtensorformer` container package.

### INFERENCE
1. The model architecture that produced `qtensorformer_small.pt` was developed in an ephemeral Docker environment and was never added to the git repository source tree.
2. Any attempt to load `qtensorformer_small.pt` into any repository class would require `strict=False`, which would silently discard all 113 trained weight tensors and run random uninitialized weights, violating scientific integrity.

### UNRESOLVED
- The original source repository of the `/app/qtensorformer` package that created `qtensorformer_small.pt` cannot be retrieved from git history.

---

## 3. Checkpoint Forensic Specification

| Attribute | Checkpoint Value |
| :--- | :--- |
| **File Name** | `qtensorformer_small.pt` |
| **File Size** | 5,432,305 bytes |
| **Object Type** | `dict` |
| **Top-Level Keys** | `['model_state_dict', 'config', 'params', 'baseline_params', 'compression_ratio', 'version']` |
| **Parameter Count** | 1,347,747 |
| **Tensor Count** | 113 |
| **Vocabulary Dimension** | 10,000 (`embeddings.token_embedding.weight`: `[10000, 128]`) |
| **Hidden Dimension ($d_{\text{model}}$)** | 128 |
| **Number of Layers** | 3 (`blocks.0`, `blocks.1`, `blocks.2`) |
| **Attention Projection** | `q_proj`: `[4, 32]`, `classical_proj`: `[128, 128]` |
| **FFN Factorization** | TensorLy TT Factorized (`factor_0..5` for `ffn` and `ffn2`) |
| **Language Model Head** | TensorLy TT Factorized (`lm_head.layer.weight.factors.factor_0..5`) |
| **Positional Encoding Length** | 64 (`embeddings.position_embedding.weight`: `[64, 128]`) |

---

## 4. Candidate Model Load Matrix

| Candidate Model | Config Class Used | Config Construction | Model Construction | `strict=True` Load | Common Keys | Missing Keys | Unexpected Keys | Shape Mismatches | Param Count Match |
| :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
| **A1: `q_tensor_former.QTensorFormer` (Default)** | `q_tensor_former.CFG()` | **PASS** | **PASS** | **FAIL** | 0 | 69 | 113 | N/A (0 key overlap) | 115,428 vs 1,347,747 (No) |
| **A2: `q_tensor_former.QTensorFormer` (Matched Dims)** | `q_tensor_former.CFG(vocab_size=10000, d_model=128, n_layers=3, max_seq_len=64)` | **PASS** | **PASS** | **FAIL** | 0 | 101 | 113 | N/A (0 key overlap) | 1,347,747 vs 1,347,747 (No) |
| **B1: `src.models.QTensorFormer` (Default)** | `src.config.ModelConfig()` | **PASS** | **PASS** | **FAIL** | 0 | 125 | 113 | N/A (0 key overlap) | 1,460,860 vs 1,347,747 (No) |
| **B2: `src.models.QTensorFormer` (Matched Dims)** | `src.config.ModelConfig(vocab_size=10000, d_model=128, n_layers=3, max_seq_len=64)` | **PASS** | **PASS** | **FAIL** | 0 | 185 | 113 | N/A (0 key overlap) | 1,526,428 vs 1,347,747 (No) |
| **C1: `q_tensor_former_v2.QTensorFormer` (Default)** | `q_tensor_former_v2.Config()` | **PASS** | **PASS** | **FAIL** | 0 | 73 | 113 | N/A (0 key overlap) | 1,476,004 vs 1,347,747 (No) |
| **C2: `q_tensor_former_v2.QTensorFormer` (Matched Dims)** | `q_tensor_former_v2.Config(vocab_size=10000, d_model=128, n_layers=3, max_seq_len=64)` | **PASS** | **PASS** | **FAIL** | 0 | 107 | 113 | N/A (0 key overlap) | 1,542,044 vs 1,347,747 (No) |
| **D1: `src.hf_model.QTensorFormerForCausalLM`** | `src.hf_model.QTensorFormerConfig()` | **PASS** | **PASS** | **FAIL** | 0 | 125 | 113 | N/A (0 key overlap) | 1,460,860 vs 1,347,747 (No) |

---

## 5. Failure Diagnosis & Scientific Conclusion

1. **Exact Error Raised**:
   ```text
   RuntimeError: Error(s) in loading state_dict for QTensorFormer:
       Missing key(s) in state_dict: ...
       Unexpected key(s) in state_dict: "embeddings.token_embedding.weight", ...
   ```
2. **Scientific Invariant**:
   Under the rule `NO RESULT > INVALID RESULT`, loading weights with `strict=False` or synthesizing dummy modules is strictly forbidden.
3. **Verdict**:
   `CHECKPOINT_STRICT_LOAD: FAIL` for all candidate models in the repository.