import json import os import platform import torch import transformers from huggingface_hub import hf_hub_download from transformers import AutoTokenizer from remedi import RemeDiUPMModelLM from remedi.configuration_remedi import RemeDiConfig from engine import MODEL_ID, MODEL_REVISION, TOKENIZER_ID TOKENIZER_REVISION = "08b83a6feb34df1a6011b80c3c00c7563e963b07" def validate_tokenizer(tokenizer, config): # Upstream stores <|mdm_mask|> in the vocabulary but leaves mask_token unset. if tokenizer.convert_tokens_to_ids("<|mdm_mask|>") != config.mask_token_id: raise RuntimeError("Base tokenizer mask vocabulary ID does not match checkpoint") tokenizer.mask_token = "<|mdm_mask|>" if tokenizer.mask_token_id != config.mask_token_id or tokenizer.eos_token_id != config.eos_token_id: raise RuntimeError("Base tokenizer special IDs do not match checkpoint") if len(tokenizer) > config.vocab_size: raise RuntimeError("Tokenizer exceeds the model vocabulary") def load_runtime(): """Called at module startup, outside the ZeroGPU lease (CUDA emulation).""" token = os.environ.get("HF_TOKEN") or None path = hf_hub_download(MODEL_ID, "config.json", revision=MODEL_REVISION, token=token) with open(path) as f: raw = json.load(f) # The checkpoint's auto_map points at absent LLaDA files. Use the vendored class. raw.pop("auto_map", None) raw.pop("model_type", None) config = RemeDiConfig(**raw) tokenizer = AutoTokenizer.from_pretrained(TOKENIZER_ID, revision=TOKENIZER_REVISION, token=token, trust_remote_code=False) validate_tokenizer(tokenizer, config) model, loading = RemeDiUPMModelLM.from_pretrained( MODEL_ID, revision=MODEL_REVISION, config=config, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, output_loading_info=True, token=token, ) problems = {k: loading.get(k) for k in ("missing_keys", "unexpected_keys", "mismatched_keys", "error_msgs") if loading.get(k)} if problems: raise RuntimeError(f"Checkpoint does not strictly match vendored architecture: {problems}") model.eval().requires_grad_(False).to("cuda") # The inference-only policy path need not invoke activation checkpointing. model.model._activation_checkpoint_fn = lambda fn, *args, **kwargs: fn(*args, **kwargs) diagnostics = {"model": MODEL_ID, "model_revision": MODEL_REVISION, "tokenizer": TOKENIZER_ID, "tokenizer_revision": TOKENIZER_REVISION, "upstream_code_revision": "2e1b19d34db09e345898f6c5b6ed60418659821a", "parameters": sum(p.numel() for p in model.parameters()), "dtype": str(next(model.parameters()).dtype), "python": platform.python_version(), "torch": torch.__version__, "transformers": transformers.__version__, "strict_loading": "passed", "special_ids": "passed"} return model, tokenizer, diagnostics