Datasets:

myx160
/

rl-post-train-artifacts

Tags:

License:

Dataset card Files Files and versions

xet

Community

rl-post-train-artifacts / scripts /create_random_gemma3_smoke.py

myx160

Add files using upload-large-folder tool

e588892 verified 21 days ago

Raw History Blame Contribute Delete

2.44 kB

#!/usr/bin/env python
"""Create a tiny random Gemma3-text checkpoint for wiring-only smoke tests.
The public Gemma3-270M checkpoint is gated.  This artifact is intentionally
*not* a pretrained model and must never be used for a result; it only lets the
ValueCriticModel forward/backward code be tested without credentials.
"""

from __future__ import annotations

import argparse
from pathlib import Path

import torch
from transformers import Gemma3ForCausalLM, Gemma3TextConfig


def main() -> int:
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--output",
        type=Path,
        default=Path("/data/myx/RL_post_train/checkpoints/pretrained/gemma-3-270m-smoke-random"),
    )
    args = parser.parse_args()
    if args.output.exists() and any(args.output.iterdir()):
        raise FileExistsError(f"refusing to overwrite non-empty directory: {args.output}")
    torch.manual_seed(0)
    config = Gemma3TextConfig(
        # The official Gemma tokenizer used by the SigLIP/Gemma checkpoint
        # has a 256k vocabulary.  Keeping the same size is important for a
        # wiring smoke test: a smaller synthetic vocabulary would trigger an
        # embedding index assertion before the model can be exercised.
        vocab_size=262144,
        hidden_size=640,
        intermediate_size=1280,
        # The value expert currently has four decoder layers and consumes the
        # corresponding first four entries of the backbone DynamicCache.
        # Keep at least that many layers in the synthetic backbone so cache
        # indexing exercises the real two-stage path.
        num_hidden_layers=4,
        # Gemma-3 270M uses grouped-query attention with one KV head.  The
        # RECAP value experts also use one KV head, so this must match for the
        # two-stage DynamicCache hand-off to be shape-compatible.
        num_attention_heads=4,
        num_key_value_heads=1,
        head_dim=256,
        max_position_embeddings=2048,
        layer_types=["full_attention"] * 4,
        tie_word_embeddings=True,
    )
    model = Gemma3ForCausalLM(config)
    args.output.mkdir(parents=True, exist_ok=True)
    model.save_pretrained(args.output, safe_serialization=True)
    print(
        f"created random Gemma3-text smoke checkpoint: {args.output} "
        f"params={sum(p.numel() for p in model.parameters())}"
    )
    return 0


if __name__ == "__main__":
    raise SystemExit(main())

	#!/usr/bin/env python
	"""Create a tiny random Gemma3-text checkpoint for wiring-only smoke tests.

	The public Gemma3-270M checkpoint is gated. This artifact is intentionally
	not a pretrained model and must never be used for a result; it only lets the
	ValueCriticModel forward/backward code be tested without credentials.
	"""

	from __future__ import annotations

	import argparse
	from pathlib import Path

	import torch
	from transformers import Gemma3ForCausalLM, Gemma3TextConfig


	def main() -> int:
	parser = argparse.ArgumentParser()
	parser.add_argument(
	"--output",
	type=Path,
	default=Path("/data/myx/RL_post_train/checkpoints/pretrained/gemma-3-270m-smoke-random"),
	)
	args = parser.parse_args()
	if args.output.exists() and any(args.output.iterdir()):
	raise FileExistsError(f"refusing to overwrite non-empty directory: {args.output}")
	torch.manual_seed(0)
	config = Gemma3TextConfig(
	# The official Gemma tokenizer used by the SigLIP/Gemma checkpoint
	# has a 256k vocabulary. Keeping the same size is important for a
	# wiring smoke test: a smaller synthetic vocabulary would trigger an
	# embedding index assertion before the model can be exercised.
	vocab_size=262144,
	hidden_size=640,
	intermediate_size=1280,
	# The value expert currently has four decoder layers and consumes the
	# corresponding first four entries of the backbone DynamicCache.
	# Keep at least that many layers in the synthetic backbone so cache
	# indexing exercises the real two-stage path.
	num_hidden_layers=4,
	# Gemma-3 270M uses grouped-query attention with one KV head. The
	# RECAP value experts also use one KV head, so this must match for the
	# two-stage DynamicCache hand-off to be shape-compatible.
	num_attention_heads=4,
	num_key_value_heads=1,
	head_dim=256,
	max_position_embeddings=2048,
	layer_types=["full_attention"] * 4,
	tie_word_embeddings=True,
	)
	model = Gemma3ForCausalLM(config)
	args.output.mkdir(parents=True, exist_ok=True)
	model.save_pretrained(args.output, safe_serialization=True)
	print(
	f"created random Gemma3-text smoke checkpoint: {args.output} "
	f"params={sum(p.numel() for p in model.parameters())}"
	)
	return 0


	if __name__ == "__main__":
	raise SystemExit(main())