()
| 99 | |
| 100 | |
| 101 | def main(): |
| 102 | args = parse_args() |
| 103 | torch.manual_seed(args.seed) |
| 104 | dtype = _DTYPES[args.dtype] |
| 105 | |
| 106 | print(f"Loading {args.model_path} (dtype={args.dtype}, device_map={args.device_map}) ...") |
| 107 | model = AutoModelForCausalLM.from_pretrained( |
| 108 | args.model_path, dtype=dtype, device_map=args.device_map, trust_remote_code=False, |
| 109 | ) |
| 110 | model.eval() |
| 111 | config = model.config |
| 112 | layout = GqlaLayout.from_config(config, args.num_kv_heads) |
| 113 | print( |
| 114 | f"H={layout.num_heads} heads -> G={layout.num_kv_heads} KV groups " |
| 115 | f"(group_size={layout.group_size}); KV cache reduction = {layout.group_size}x" |
| 116 | ) |
| 117 | |
| 118 | tokenizer = AutoTokenizer.from_pretrained(args.model_path, trust_remote_code=False) |
| 119 | |
| 120 | ppl_loader = ppl_orig = ppl_gqla = None |
| 121 | if args.eval_ppl: |
| 122 | print(f"Building PPL loader ({args.eval_ppl_dataset}, seqlen={args.eval_ppl_seqlen}) ...") |
| 123 | ppl_loader = prepare_ppl_dataloader( |
| 124 | tokenizer, args.eval_ppl_dataset, args.eval_ppl_seqlen, |
| 125 | args.eval_ppl_batch_size, max_chunks=args.eval_ppl_max_chunks, |
| 126 | ) |
| 127 | print("Eval PPL on original MLA model ...") |
| 128 | ppl_orig = evaluate_ppl(model, ppl_loader) |
| 129 | print(f" Original MLA PPL = {ppl_orig:.4f}") |
| 130 | |
| 131 | print( |
| 132 | f"\nBuilding calibration ({args.cal_dataset}, n={args.cal_nsamples}, " |
| 133 | f"len={args.cal_seqlen}) ..." |
| 134 | ) |
| 135 | batches = prepare_calibration_inputs( |
| 136 | tokenizer, args.cal_dataset, args.cal_nsamples, args.cal_seqlen, seed=args.seed, |
| 137 | ) |
| 138 | |
| 139 | # First-rank device for input_ids; HF device_map auto-routes through the rest. |
| 140 | input_device = next(model.parameters()).device |
| 141 | print(f"Gathering per-layer hidden states ({config.num_hidden_layers} layers) ...") |
| 142 | if args.hessian_pca: |
| 143 | calib, final_hiddens = gather_calibration_hidden_states( |
| 144 | model, batches, input_device, capture_final=True, |
| 145 | ) |
| 146 | else: |
| 147 | calib = gather_calibration_hidden_states(model, batches, input_device) |
| 148 | final_hiddens = None |
| 149 | |
| 150 | # Hessian-aware token weights (NLL of next token from the teacher's lm_head). |
| 151 | # Computed once on the cached final hidden states; threaded into every layer's |
| 152 | # cov accumulation. When --hessian_pca is OFF, token_weights stays None and the |
| 153 | # cov path is byte-identical to the legacy code. |
| 154 | token_weights = None |
| 155 | if args.hessian_pca: |
| 156 | print(f"[hessian_pca] computing per-token NLL weights ({args.hessian_mode}) ...") |
| 157 | token_weights = compute_token_weights_nll(model, final_hiddens, batches) |
| 158 | stats = diagnose_weights(token_weights) |
no test coverage detected