tarekziade HF Staff
Publish MIC usage, slim fixture and reviewed Transformers expectations
43a3e9a verified |
Download integrations/transformers/usage.md from huggingface-mic/recurrentgemma-2b-mic: direct link, hf CLI and curl.
- Browser
- Download file 5.03 kB
-
https://huggingface.co/huggingface-mic/recurrentgemma-2b-mic/resolve/main/integrations/transformers/usage.md
- Command line
-
hf download hf://huggingface-mic/recurrentgemma-2b-mic/integrations/transformers/usage.md
-
curl -L -o usage.md https://huggingface.co/huggingface-mic/recurrentgemma-2b-mic/resolve/main/integrations/transformers/usage.md
5.03 kB
| # RecurrentGemma: causal language modeling with Transformers | |
| This page is the Transformers contract of the RecurrentGemma MIC and also a usage example. [doc-builder](https://github.com/huggingface/doc-builder/blob/main/docs/runnable-code-blocks.md) renders the `runnable` blocks below as documentation, and its pytest plugin runs them as written against `google/recurrentgemma-2b`. The MIC runner runs the same blocks against each slim fixture and each checkpoint in [`integration.json`](../../integration.json). Lines marked `# doc-builder: hide` only matter there: they swap in the pinned checkpoint or variant and print the result record. | |
| RecurrentGemma mixes recurrent (RG-LRU + conv1d) blocks with sliding-window attention blocks, so it is a decoder-only causal LM with the usual `input_ids` / `attention_mask` / `labels` interface. The checkpoint is a base (pretraining) model: it has a tokenizer and a causal-LM head, but no chat template. | |
| ## Next-token logits for a padded batch | |
| ```py runnable:causal_lm | |
| import torch | |
| from transformers import AutoModelForCausalLM, AutoTokenizer, set_seed | |
| import json # doc-builder: hide | |
| import os # doc-builder: hide | |
| model_id = "google/recurrentgemma-2b" | |
| model_id = os.environ.get("MIC_MODEL", model_id) # doc-builder: hide | |
| variant = json.loads(os.environ.get("MIC_MODEL_KWARGS", "{}")) # doc-builder: hide | |
| device = "cuda" if torch.cuda.is_available() else "cpu" | |
| set_seed(0) # seeds torch, NumPy, and Python: loading and any dropout draw random numbers | |
| tokenizer = AutoTokenizer.from_pretrained(model_id) | |
| model = AutoModelForCausalLM.from_pretrained(model_id).to(device) | |
| if variant: # doc-builder: hide | |
| del model # doc-builder: hide | |
| model = AutoModelForCausalLM.from_pretrained(model_id, **variant).to(device) # doc-builder: hide | |
| model.eval() | |
| # Two prompts of different lengths so padding and the attention mask are exercised. | |
| prompts = [ | |
| "The capital of France is", | |
| "Two plus two equals", | |
| ] | |
| # Left padding keeps the last position a real token for every row. | |
| tokenizer.pad_token = tokenizer.eos_token | |
| tokenizer.padding_side = "left" | |
| inputs = tokenizer(prompts, return_tensors="pt", padding=True).to(device) | |
| # The recurrent blocks reset their state where position_ids == 0 and do not read the mask, | |
| # so derive positions from it (as generate does): the first real token of each row is 0. | |
| inputs["position_ids"] = (inputs["attention_mask"].cumsum(-1) - 1).clamp(min=0) | |
| with torch.no_grad(): | |
| logits = model(**inputs).logits | |
| # Probe a few logits: the last position of each row, and the first real token of row 0. | |
| last = logits[:, -1, :] | |
| first_real = logits[0, inputs["attention_mask"][0].bool().nonzero()[0].item(), :] | |
| probe_ids = torch.tensor([tokenizer.convert_tokens_to_ids("Paris"), tokenizer.convert_tokens_to_ids("4")], device=device) | |
| print("last-position logits for 'Paris' and '4':", last[:, probe_ids].float().tolist()) | |
| ``` | |
| ## One training step | |
| ```py runnable:causal_lm:2 | |
| # Causal-LM labels: copy the input ids and mask padded positions out of the loss. | |
| labels = inputs["input_ids"].clone() | |
| labels[inputs["attention_mask"] == 0] = -100 | |
| model.train() | |
| # Full fine-tuning of 2.7B float32 parameters does not fit a 24 GB GPU: train the | |
| # (tied) token embeddings and the first recurrent block, and freeze the rest. | |
| for name, param in model.named_parameters(): | |
| param.requires_grad = name.startswith(("model.embed_tokens.", "model.layers.0.")) | |
| optimizer = torch.optim.SGD([p for p in model.parameters() if p.requires_grad], lr=0.01) | |
| loss = model(**inputs, labels=labels).loss | |
| loss.backward() | |
| before = {n: p.detach().cpu().clone() for n, p in model.named_parameters() if p.requires_grad} # doc-builder: hide | |
| optimizer.step() | |
| print(f"loss: {loss.item():.4f}") | |
| if os.environ.get("MIC_RUNNER"): # doc-builder: hide | |
| named = dict(model.named_parameters()) | |
| checked = ( | |
| "model.embed_tokens.weight", | |
| "model.layers.0.temporal_block.rg_lru.recurrent_param", | |
| "model.layers.0.mlp_block.gate_proj.weight", | |
| "model.layers.0.temporal_block.linear_x.weight", | |
| ) | |
| result = { | |
| "runtime": { | |
| "device": model.device.type, | |
| "dtype": str(model.dtype).removeprefix("torch."), | |
| "attn_implementation": model.config._attn_implementation, | |
| }, | |
| "logits": { | |
| "last": last[:, probe_ids].float().tolist(), | |
| "first_real": first_real[probe_ids].float().tolist(), | |
| }, | |
| "training": { | |
| "loss": loss.item(), | |
| "gradients": { | |
| n: bool( | |
| named[n].grad is not None | |
| and torch.isfinite(named[n].grad).all() | |
| and torch.count_nonzero(named[n].grad) > 0 | |
| ) | |
| for n in checked | |
| }, | |
| "updated": any( | |
| not torch.equal(p.detach().cpu(), before[n]) | |
| for n, p in model.named_parameters() | |
| if n in before | |
| ), | |
| }, | |
| } | |
| print(json.dumps(result)) | |
| ``` | |