| import os, json, glob, gc |
| import torch |
| from safetensors.torch import load_file |
| from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer |
|
|
| CACHE = "/root/crossarch-accuracy/cache/ckpt" |
|
|
| def meta_dir(fam): return f"{CACHE}/{fam}_meta" |
| def sd_path(fam, step): return f"{CACHE}/{fam}_{step}.safetensors" |
|
|
| def load_sd(fam, step, device="cpu", dtype=torch.float32): |
| return {k: v.to(device, dtype) for k, v in load_file(sd_path(fam, step)).items()} |
|
|
| def build_model(fam, device): |
| cfg = AutoConfig.from_pretrained(meta_dir(fam)) |
| m = AutoModelForCausalLM.from_config(cfg, dtype=torch.float16) |
| m = m.to(device).eval() |
| tok = AutoTokenizer.from_pretrained(meta_dir(fam)) |
| if tok.pad_token is None: tok.pad_token = tok.eos_token |
| return m, tok |
|
|
| ALIAS = {"embed_out.weight": "lm_head.weight", "lm_head.weight": "embed_out.weight"} |
|
|
| @torch.no_grad() |
| def install(model, sd): |
| """Copy a state dict into a live model, tolerating the transformers v4/v5 head rename.""" |
| params = dict(model.named_parameters()); bufs = dict(model.named_buffers()) |
| miss = [] |
| for k, v in sd.items(): |
| cands = [k, ALIAS.get(k), |
| k.replace("gpt_neox.", "") if k.startswith("gpt_neox.") else "gpt_neox." + k] |
| t = None |
| for c in cands: |
| if c is None: continue |
| t = params.get(c, bufs.get(c)) |
| if t is not None: break |
| if t is None or t.shape != v.shape: |
| miss.append(k); continue |
| t.copy_(v.to(t.dtype)) |
| return miss |
|
|
