hanzo/ml
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

release: ml 0.11.8 — fused MoE expert-combine + f32-direct quantize (MoE prefill at llama parity)

Browse Source 
moe_combine_{f16,bf16,f32} replaces the broadcast_mul+sum expert-combine (fast_sum 487->8ms);
f32-direct activation quantize kills a 178ms cast. Qwen3-30B-A3B-Q4_K_M prefill 939->1167 T/s
(+19.5% isolated A/B, 1.11x of llama.cpp-HIP on the same gfx1151), decode flat. nbad=0
(moe_combine_numeric f16/bf16/f32 + existing MoE oracles). Stacks on the 0.11.7 fused GEMM.
This commit is contained in:
hanzo-dev
2026-06-25 22:10:05 +00:00
parent 5ff6df6b5a
commit ad531713d0
27 changed files with 43 additions and 43 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Cargo.toml
+12 -12
View File
@@ -33,7 +33,7 @@ exclude = [
resolver = "2"
[workspace.package]
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
@@ -43,23 +43,23 @@ license = "MIT OR Apache-2.0"
[workspace.dependencies]
rocm-rs = { package = "hanzo-rocm", version = "0.5.2", default-features = false, features = ["macros"] }
hanzo-rocm-kernels = { path = "./hanzo-rocm-kernels", version = "0.11.7" }
hanzo-rocm-kernels = { path = "./hanzo-rocm-kernels", version = "0.11.8" }
ab_glyph = "0.2.23"
accelerate-src = { version = "0.3.2" }
anyhow = { version = "1", features = ["backtrace"] }
byteorder = "1.4.3"
chrono = "0.4"
# Hanzo-renamed workspace dependencies (at version 0.10.x)
hanzo-ml = { path = "./hanzo-ml", package = "hanzo-ml", version = "0.11.7" }
hanzo-datasets = { path = "./hanzo-datasets", version = "0.11.7" }
hanzo-flash-attn = { path = "./hanzo-flash-attn", version = "0.11.7" }
hanzo-flash-attn-v3 = { path = "./hanzo-flash-attn-v3", version = "0.11.7" }
hanzo-kernels = { path = "./hanzo-kernels", version = "0.11.7" }
hanzo-metal-kernels = { path = "./hanzo-metal-kernels", version = "0.11.7" }
hanzo-nn = { path = "./hanzo-nn", version = "0.11.7" }
hanzo-onnx = { path = "./hanzo-onnx", version = "0.11.7" }
hanzo-transformers = { path = "./hanzo-transformers", version = "0.11.7" }
hanzo-ug = { path = "./hanzo-ug", version = "0.11.7" }
hanzo-ml = { path = "./hanzo-ml", package = "hanzo-ml", version = "0.11.8" }
hanzo-datasets = { path = "./hanzo-datasets", version = "0.11.8" }
hanzo-flash-attn = { path = "./hanzo-flash-attn", version = "0.11.8" }
hanzo-flash-attn-v3 = { path = "./hanzo-flash-attn-v3", version = "0.11.8" }
hanzo-kernels = { path = "./hanzo-kernels", version = "0.11.8" }
hanzo-metal-kernels = { path = "./hanzo-metal-kernels", version = "0.11.8" }
hanzo-nn = { path = "./hanzo-nn", version = "0.11.8" }
hanzo-onnx = { path = "./hanzo-onnx", version = "0.11.8" }
hanzo-transformers = { path = "./hanzo-transformers", version = "0.11.8" }
hanzo-ug = { path = "./hanzo-ug", version = "0.11.8" }
clap = { version = "4.2.4", features = ["derive"] }
criterion = { version = "0.8", default-features = false }
cudarc = { version = "0.19.4", features = [
hanzo-datasets/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-datasets"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-flash-attn-build/Cargo.toml
+1 -1
View File
@@ -7,7 +7,7 @@
[package]
name = "hanzo-flash-attn-build"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Build-time helper that fetches and resolves the pinned NVIDIA CUTLASS include directory for the hanzo flash-attention CUDA kernels."
repository = "https://github.com/hanzoai/ml"
hanzo-flash-attn-v3/Cargo.toml
+2 -2
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-flash-attn-v3"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "FlashAttention-3 (Hopper/sm90) layer for the Hanzo ML framework."
@@ -12,7 +12,7 @@ readme = "README.md"
exclude = ["cutlass/docs/**", "cutlass/test/**", "cutlass/examples/**", "cutlass/tools/**", "cutlass/media/**"]
[dependencies]
hanzo-ml = { path = "../hanzo-ml", features = ["cuda"], package = "hanzo-ml", version = "0.11.7" }
hanzo-ml = { path = "../hanzo-ml", features = ["cuda"], package = "hanzo-ml", version = "0.11.8" }
half = { version = "2.3.1", features = ["num-traits"] }
[build-dependencies]
hanzo-flash-attn/Cargo.toml
+3 -3
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-flash-attn"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
@@ -10,13 +10,13 @@ license = "BSD-3-Clause OR Apache-2.0"
readme = "README.md"
[dependencies]
hanzo-ml = { path = "../hanzo-ml", features = ["cuda"], package = "hanzo-ml", version = "0.11.7" }
hanzo-ml = { path = "../hanzo-ml", features = ["cuda"], package = "hanzo-ml", version = "0.11.8" }
half = { version = "2.3.1", features = ["num-traits"] }
[build-dependencies]
bindgen_cuda = "0.1.5"
anyhow = { version = "1", features = ["backtrace"] }
hanzo-flash-attn-build = { path = "../hanzo-flash-attn-build", version = "0.11.7" }
hanzo-flash-attn-build = { path = "../hanzo-flash-attn-build", version = "0.11.8" }
[dev-dependencies]
anyhow = { version = "1", features = ["backtrace"] }
hanzo-kernels/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-kernels"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-metal-kernels/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-metal-kernels"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-ml-book/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-ml-book"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-ml-examples/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-ml-examples"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-ml-pyo3/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-ml-pyo3"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-ml-wasm-examples/bert/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-ml-wasm-example-bert"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-ml-wasm-examples/blip/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-ml-wasm-example-blip"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-ml-wasm-examples/llama2-c/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-ml-wasm-example-llama2"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-ml-wasm-examples/moondream/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-ml-wasm-example-moondream"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-ml-wasm-examples/phi/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-ml-wasm-example-phi"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-ml-wasm-examples/segment-anything/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-ml-wasm-example-sam"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-ml-wasm-examples/t5/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-ml-wasm-example-t5"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-ml-wasm-examples/whisper/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-ml-wasm-example-whisper"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-ml-wasm-examples/yolo/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-ml-wasm-example-yolo"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-ml-wasm-tests/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-ml-wasm-tests"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "WASM tests for hanzo"
keywords = ["blas", "tensor", "machine-learning"]
hanzo-ml/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-ml"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-nn/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-nn"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-onnx/Cargo.toml
+3 -3
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-onnx"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
@@ -9,8 +9,8 @@ categories = ["science"]
license = "BSD-3-Clause OR Apache-2.0"
[dependencies]
hanzo-ml = { path = "../hanzo-ml", package = "hanzo-ml", version = "0.11.7" }
hanzo-nn = { path = "../hanzo-nn", version = "0.11.7" }
hanzo-ml = { path = "../hanzo-ml", package = "hanzo-ml", version = "0.11.8" }
hanzo-nn = { path = "../hanzo-nn", version = "0.11.8" }
prost = "0.14.1"
[build-dependencies]
hanzo-rocm-kernels/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-rocm-kernels"
version = "0.11.7"
version = "0.11.8"
license = "MIT OR Apache-2.0"
edition = "2021"
hanzo-training/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-training"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Training framework for Hanzo ML models on zen-agentic-dataset"
repository = "https://github.com/hanzoai/ml"
hanzo-transformers/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-transformers"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"
hanzo-ug/Cargo.toml
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "hanzo-ug"
version = "0.11.7"
version = "0.11.8"
edition = "2021"
description = "Minimalist ML framework."
repository = "https://github.com/hanzoai/ml"