Qwen2¶

Weights: pretrained Keras weights live on Hugging Face under zeromodels/<variant> (each repo carries kf_config.json plus the Keras weights: model.weights.h5, or a sharded model.weights.json + shards for the larger checkpoints). Load with from_weights("zeromodels/<variant>").

Alibaba's Qwen2 dense decoder-only LLM, ported to pure Keras 3. A pre-norm transformer with RMSNorm, SwiGLU MLP, rotary embeddings, grouped-query attention and biased QKV projections. The smaller variants tie the embedding matrix to the LM head.

Links:

HF collection: Qwen2
Paper: Qwen2 Technical Report (arXiv:2407.10671)
HF docs: transformers/model_doc/qwen2

Variants¶

Load any of these with from_weights("zeromodels/<variant>"). The -instruct suffix marks instruction-tuned checkpoints (use the chat template); bare names are base models.

Variant	Hub
`qwen2-0.5b`	`zeromodels/qwen2-0.5b`
`qwen2-0.5b-instruct`	`zeromodels/qwen2-0.5b-instruct`
`qwen2-1.5b`	`zeromodels/qwen2-1.5b`
`qwen2-1.5b-instruct`	`zeromodels/qwen2-1.5b-instruct`
`qwen2-7b`	`zeromodels/qwen2-7b`
`qwen2-7b-instruct`	`zeromodels/qwen2-7b-instruct`
`qwen2-72b`	`zeromodels/qwen2-72b`
`qwen2-72b-instruct`	`zeromodels/qwen2-72b-instruct`

API¶

`Qwen2Model`¶

The decoder backbone, no LM head. Returns {"last_hidden_state": (batch, seq, embed_dim)}.

Arg	Default	Meaning
`vocab_size`	`151936`	token vocabulary size
`embed_dim`	`896`	model width
`mlp_dim`	`4864`	MLP inner width
`num_layers`	`24`	decoder blocks
`num_heads`	`14`	query heads
`num_kv_heads`	`2`	key/value heads (GQA)
`head_dim`	`None`	per-head width
`norm_eps`	`1e-06`	RMSNorm epsilon
`rope_theta`	`1000000.0`	rotary base frequency
`tie_embeddings`	`True`	reuse the embedding matrix as the LM head

`Qwen2TextGenerate`¶

Qwen2Model plus a (tied) LM head. Returns {"logits": (batch, seq, vocab_size)} and adds .generate(). Same constructor arguments as Qwen2Model.

generate(
    input_ids,
    attention_mask=None,
    max_new_tokens=None,
    eos_token_id=None,
    sampler=None,
    seed=None,
    **prefill_inputs,
)

Arg	Default	Meaning
`input_ids`	required	`(batch, seq)` token ids
`attention_mask`	`None`	`(batch, seq)` 1 = keep, 0 = padding
`max_new_tokens`	`None`	tokens to generate
`eos_token_id`	`None`	stop token (defaults to the tokenizer's)
`sampler`	`None`	sampling strategy; greedy when unset
`seed`	`None`	seed for stochastic samplers

`Qwen2Tokenizer`¶

Tokenizer on the tokenizers backend.

Qwen2Tokenizer(hf_id=None, tokenizer_file=None)

Arg	Default	Meaning
`hf_id`	`None`	Hub repo to pull the tokenizer files from
`tokenizer_file`	`None`	explicit path to a `tokenizer.json`

Calling it returns {"input_ids", "attention_mask"}, padded across the batch. It accepts a plain string, a list of strings (a batch), or a chat-message list, which is routed through apply_chat_template automatically. Decode with .decode(ids) for one sequence or .batch_decode(ids) for a batch.

End-to-end example¶

Single input¶

import os

os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from zeromodels.models.qwen2 import Qwen2TextGenerate, Qwen2Tokenizer

model = Qwen2TextGenerate.from_weights("zeromodels/qwen2-0.5b-instruct")
tokenizer = Qwen2Tokenizer.from_weights("zeromodels/qwen2-0.5b-instruct")

inputs = tokenizer(
    [{"role": "user", "content": "Explain rotary embeddings in one sentence."}]
)
outputs = model.generate(**inputs, max_new_tokens=64)

print(tokenizer.decode(outputs[0]))

Batch¶

Pass a list of strings. The tokenizer pads them and generate runs the batch together:

prompts = [
    "The capital of France is",
    "In one sentence, what is a transformer?",
    "Write a haiku about GPUs.",
]
inputs = tokenizer(prompts)  # {"input_ids": (3, seq), "attention_mask": (3, seq)}
outputs = model.generate(**inputs, max_new_tokens=64)

for text in tokenizer.batch_decode(outputs):
    print(text)

Backbone only¶

from zeromodels.models.qwen2 import Qwen2Model

backbone = Qwen2Model.from_weights("zeromodels/qwen2-0.5b")
hidden = backbone(inputs)["last_hidden_state"]  # (batch, seq, embed_dim)

Loading from the Hub¶

Any Hub repo with this architecture works via the hf: prefix, including community fine-tunes:

model = Qwen2TextGenerate.from_weights("hf:Qwen/Qwen2-0.5B")

Lower memory¶

The 72B checkpoint needs quantization to fit comfortably on a single 80GB GPU. See quantization.md:

model = Qwen2TextGenerate.from_weights(
    "zeromodels/qwen2-72b-instruct",
    quantization="int8",
)