| 49 | return result |
| 50 | |
| 51 | def quantize(self, parameter_group, overflow, eigenvalue_enabled, block_eigenvalue={}): |
| 52 | |
| 53 | if overflow and not eigenvalue_enabled: |
| 54 | return |
| 55 | |
| 56 | self.step() |
| 57 | |
| 58 | self.update_fp16_ratio() |
| 59 | |
| 60 | for i in range(len(parameter_group)): |
| 61 | for p in parameter_group[i]: |
| 62 | if len(p.size()) > 1 and hasattr(p, "start_bits") and p.start_bits: |
| 63 | param_id = id(p) |
| 64 | if block_eigenvalue is None: |
| 65 | eigenvalue, layer_id = None, 0 |
| 66 | else: |
| 67 | eigenvalue, layer_id = block_eigenvalue[param_id] if param_id in block_eigenvalue else (None, |
| 68 | 0) |
| 69 | if eigenvalue is not None: |
| 70 | factor = 1 + math.floor(eigenvalue * 4) |
| 71 | p.data = self.compute_quantization(p.data, layer_id, factor) |
| 72 | else: |
| 73 | p.data = self.compute_quantization(p, layer_id) |
| 74 | |
| 75 | def step(self): |
| 76 | self.qsteps += 1 |