Mixtral¶

On-the-fly conversion: these weights are not mirrored as preconverted .weights.h5 under zeromodels/. from_weights("<variant>") downloads the original safetensors from the Hub and converts them in process on every load, because checkpoints this large are impractical to re-host. Pass cache_converted=True to keep the converted result and skip the download and conversion next time. See Loading Weights.

Mistral AI's sparse Mixture-of-Experts LLM, ported to pure Keras 3. The attention stack matches Mistral; each MLP is replaced by a router over experts, of which only num_experts_per_tok run per token.

Memory is governed by total parameters, not active ones: every expert stays resident, so plan capacity for the full weight count even though a fraction runs per token.

Links:

Paper: Mixtral of Experts (arXiv:2401.04088)
HF docs: transformers/model_doc/mixtral

Variants¶

Load any of these with from_weights("<variant>").

Variant	Hub
`mixtral-8x7b`	`mistralai/Mixtral-8x7B-v0.1`
`mixtral-8x7b-instruct`	`mistralai/Mixtral-8x7B-Instruct-v0.1`
`mixtral-8x22b`	`mistralai/Mixtral-8x22B-v0.1`
`mixtral-8x22b-instruct`	`mistralai/Mixtral-8x22B-Instruct-v0.1`

API¶

`MixtralModel`¶

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

Arg	Default	Meaning
`vocab_size`	`32000`	token vocabulary size
`embed_dim`	`4096`	model width
`mlp_dim`	`14336`	MLP inner width
`num_layers`	`32`	decoder blocks
`num_heads`	`32`	query heads
`num_kv_heads`	`8`	key/value heads (GQA)
`head_dim`	`None`	per-head width
`num_experts`	`8`	expert count
`num_experts_per_tok`	`2`	experts routed per token
`norm_eps`	`1e-05`	RMSNorm epsilon
`rope_theta`	`1000000.0`	rotary base frequency
`sliding_window`	`None`	local attention span
`tie_embeddings`	`False`	reuse the embedding matrix as the LM head

`MixtralTextGenerate`¶

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

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

`MixtralTokenizer`¶

Tokenizer on the tokenizers backend.

MixtralTokenizer(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.mixtral import MixtralTextGenerate, MixtralTokenizer

model = MixtralTextGenerate.from_weights("mixtral-8x7b")
tokenizer = MixtralTokenizer.from_weights("mixtral-8x7b")

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.mixtral import MixtralModel

backbone = MixtralModel.from_weights("mixtral-8x7b")
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 = MixtralTextGenerate.from_weights("hf:mistralai/Mixtral-8x7B-v0.1")

Lower memory¶

Larger checkpoints load in bf16 or weight-only quantized. See quantization.md:

model = MixtralTextGenerate.from_weights(
    "mixtral-8x7b", quantization="int8", load_dtype="bfloat16"
)