mirror of
https://github.com/hanzoai/ml.git
synced 2026-11-20 23:08:29 +00:00
Every crate this merge changed moves one patch over what crates.io serves, and every crate it did not change holds still. hanzo-kernels now requires hanzo-bindgen-cuda 0.1.8, without which nothing built from the registry would pick up the toolkit fix and the compiler would keep disagreeing with the headers. hanzo-bindgen-cuda 0.1.8, hanzo-metal-kernels 0.11.45, hanzo-kernels 0.11.37, hanzo-ml 0.11.96, hanzo-flash-attn 0.11.36, hanzo-flash-attn-v3 0.11.10, hanzo-nn 0.11.40, hanzo-datasets 0.11.9, hanzo-onnx 0.11.9, hanzo-transformers 0.11.10, hanzo-training 0.11.9. Co-authored-by: Hanzo Dev <[email protected]>
80 lines
2.3 KiB
TOML
80 lines
2.3 KiB
TOML
[package]
|
|
name = "hanzo-training"
|
|
version = "0.11.9"
|
|
edition = "2021"
|
|
description = "Training framework for Hanzo ML models on zen-agentic-dataset"
|
|
repository = "https://github.com/hanzoai/ml"
|
|
keywords = ["machine-learning", "training", "transformer", "llm"]
|
|
categories = ["science", "algorithms"]
|
|
license = "MIT OR Apache-2.0"
|
|
readme = "README.md"
|
|
|
|
[dependencies]
|
|
hanzo-ml = { workspace = true }
|
|
hanzo-nn = { workspace = true }
|
|
hanzo-transformers = { workspace = true }
|
|
hanzo-datasets = { workspace = true }
|
|
|
|
# Core dependencies
|
|
anyhow = { workspace = true }
|
|
clap = { workspace = true }
|
|
serde = { workspace = true }
|
|
serde_json = { workspace = true }
|
|
log = { workspace = true }
|
|
thiserror = { workspace = true }
|
|
tokio = { version = "1.0", features = ["full"] }
|
|
tracing = { version = "0.1" }
|
|
tracing-subscriber = { version = "0.3", features = ["env-filter", "fmt"] }
|
|
|
|
# Configuration
|
|
chrono = { workspace = true }
|
|
serde_yaml = "0.9"
|
|
toml = "0.8"
|
|
|
|
# Training specific
|
|
rand = { workspace = true }
|
|
rayon = { workspace = true }
|
|
safetensors = { workspace = true }
|
|
|
|
# Progress and metrics
|
|
indicatif = "0.18"
|
|
# wandb = { version = "0.1", optional = true } # Commented out for now
|
|
|
|
[dev-dependencies]
|
|
tempfile = "3.0"
|
|
|
|
[features]
|
|
default = []
|
|
cuda = ["hanzo-ml/cuda", "hanzo-nn/cuda"]
|
|
metal = ["hanzo-ml/metal", "hanzo-nn/metal"]
|
|
# wandb = ["dep:wandb"] # Commented out for now
|
|
|
|
# The binary is named for its crate, not for a verb it cannot perform. `hanzo-train`
|
|
# is the `hanzo-train` crate's binary — the DSpark draft trainer that reaches
|
|
# hanzo_nn::AdamW and writes an engine-loadable checkpoint. Two `[[bin]]` targets
|
|
# with one name both uplift to `target/<profile>/hanzo-train`, so whichever linked
|
|
# last decided whether `cargo build --workspace` left you a trainer or a CLI whose
|
|
# `train` subcommand refuses (see src/model.rs::unwired). One name, one artifact.
|
|
#
|
|
# What this binary can do is READ and WRITE a training config (`validate`,
|
|
# `template`); `train` refuses until the config-driven path is wired to
|
|
# hanzo-transformers. The connected training path in this crate is `grpo`.
|
|
[[bin]]
|
|
name = "hanzo-training"
|
|
path = "src/bin/train.rs"
|
|
|
|
[[bin]]
|
|
name = "hanzo-eval"
|
|
path = "src/bin/eval.rs"
|
|
|
|
[[example]]
|
|
name = "train_llama"
|
|
path = "examples/train_llama.rs"
|
|
|
|
[[example]]
|
|
name = "train_qwen"
|
|
path = "examples/train_qwen.rs"
|
|
|
|
[[example]]
|
|
name = "grpo_toy"
|
|
path = "examples/grpo_toy.rs" |