Files
zeekay 8cfdc29b5c feat(deepseek-v4): training-cache capture + v4_cache_dump — DSpark cache generation at 111.5 tok/s
V4_CAPTURE_DIR env in quantized_deepseek4.rs: during prefill, capture the HC carrier stream-sum at layers [1,11,21,31,41] ([seq,5,4096] bf16) + the post-final-norm hidden ([seq,4096]) and write one safetensors per sample + index.jsonl. Example v4_cache_dump: JSONL corpus -> teacher-forced prefill dumps. MEASURED: 300 open-perfectblend samples, 125k tokens in 18.7min = 111.5 tok/s; hiddens verified finite, carrier rms ~860 matching the layer-trace physics. Feeds DeepSpec's convert_engine_dump.py -> dspark_v4flash_quant.py training (real-hiddens 50-step smoke: loss 3.03->2.6 decreasing).
2026-07-02 21:16:03 -07:00

289 lines
6.2 KiB
TOML

[package]
name = "hanzo"
readme = "README.md"
authors = ["Hanzo AI"]
version.workspace = true
edition.workspace = true
rust-version.workspace = true
description.workspace = true
repository.workspace = true
keywords.workspace = true
categories.workspace = true
license.workspace = true
homepage.workspace = true
documentation = "https://docs.rs/hanzo"
[dependencies]
hanzo-engine.workspace = true
hanzo-macros.workspace = true
anyhow.workspace = true
thiserror.workspace = true
tokio.workspace = true
hanzo-ml.workspace = true
hanzo-nn.workspace = true
serde_json.workspace = true
serde.workspace = true
schemars.workspace = true
image.workspace = true
indexmap.workspace = true
either.workspace = true
futures.workspace = true
reqwest.workspace = true
rand.workspace = true
clap.workspace = true
walkdir.workspace = true
tracing.workspace = true
tracing-subscriber.workspace = true
# ========== Getting Started ==========
[[example]]
name = "text_generation"
path = "examples/getting_started/text_generation/main.rs"
[[example]]
name = "dspark_qwen3"
path = "examples/dspark_qwen3/main.rs"
[[example]]
name = "streaming"
path = "examples/getting_started/streaming/main.rs"
[[example]]
name = "multimodal_basic"
path = "examples/getting_started/multimodal/main.rs"
[[example]]
name = "gguf"
path = "examples/getting_started/gguf/main.rs"
[[example]]
name = "gguf_locally"
path = "examples/getting_started/gguf_locally/main.rs"
[[example]]
name = "embedding"
path = "examples/getting_started/embedding/main.rs"
# ========== Models ==========
[[example]]
name = "text_models"
path = "examples/models/text_models/main.rs"
[[example]]
name = "multimodal_models"
path = "examples/models/multimodal_models/main.rs"
[[example]]
name = "multimodal_multiturn"
path = "examples/models/multimodal_multiturn/main.rs"
[[example]]
name = "audio"
path = "examples/models/audio/main.rs"
[[example]]
name = "multimodal"
path = "examples/models/multimodal/main.rs"
[[example]]
name = "diffusion"
path = "examples/models/diffusion/main.rs"
[[example]]
name = "speech"
path = "examples/models/speech/main.rs"
[[example]]
name = "asr"
path = "examples/models/asr/main.rs"
# ========== Quantization ==========
[[example]]
name = "isq"
path = "examples/quantization/isq/main.rs"
[[example]]
name = "imatrix"
path = "examples/quantization/imatrix/main.rs"
[[example]]
name = "uqff"
path = "examples/quantization/uqff/main.rs"
[[example]]
name = "uqff_multimodal"
path = "examples/quantization/uqff_multimodal/main.rs"
[[example]]
name = "topology"
path = "examples/quantization/topology/main.rs"
[[example]]
name = "mixture_of_quant_experts"
path = "examples/quantization/mixture_of_quant_experts/main.rs"
# ========== Advanced ==========
[[example]]
name = "tools"
path = "examples/advanced/tools/main.rs"
[[example]]
name = "agent"
path = "examples/advanced/agent/main.rs"
[[example]]
name = "agent_streaming"
path = "examples/advanced/agent_streaming/main.rs"
[[example]]
name = "web_search"
path = "examples/advanced/web_search/main.rs"
[[example]]
name = "search_callback"
path = "examples/advanced/search_callback/main.rs"
[[example]]
name = "tool_callback"
path = "examples/advanced/tool_callback/main.rs"
[[example]]
name = "mcp_client"
path = "examples/advanced/mcp_client/main.rs"
[[example]]
name = "grammar"
path = "examples/advanced/grammar/main.rs"
[[example]]
name = "json_schema"
path = "examples/advanced/json_schema/main.rs"
[[example]]
name = "llguidance"
path = "examples/advanced/llguidance/main.rs"
[[example]]
name = "logits_processor"
path = "examples/advanced/logits_processor/main.rs"
[[example]]
name = "file_logging"
path = "examples/advanced/file_logging/main.rs"
[[example]]
name = "batching"
path = "examples/advanced/batching/main.rs"
[[example]]
name = "batching_embeddings"
path = "examples/advanced/batching_embeddings/main.rs"
[[example]]
name = "paged_attn"
path = "examples/advanced/paged_attn/main.rs"
[[example]]
name = "auto_device_map"
path = "examples/advanced/auto_device_map/main.rs"
[[example]]
name = "multi_model"
path = "examples/advanced/multi_model/main.rs"
[[example]]
name = "lora"
path = "examples/advanced/lora/main.rs"
[[example]]
name = "xlora"
path = "examples/advanced/xlora/main.rs"
[[example]]
name = "anymoe"
path = "examples/advanced/anymoe/main.rs"
[[example]]
name = "anymoe_lora"
path = "examples/advanced/anymoe_lora/main.rs"
[[example]]
name = "error_handling"
path = "examples/advanced/error_handling/main.rs"
[[example]]
name = "perplexity"
path = "examples/advanced/perplexity/main.rs"
[[example]]
name = "embeddings"
path = "examples/advanced/embeddings/main.rs"
[[example]]
name = "code_execution"
path = "examples/advanced/code_execution/main.rs"
[[example]]
name = "code_execution_approval"
path = "examples/advanced/code_execution_approval/main.rs"
[[example]]
name = "code_execution_files"
path = "examples/advanced/code_execution_files/main.rs"
# ========== Cookbook ==========
[[example]]
name = "cookbook_rag"
path = "examples/cookbook/rag/main.rs"
[[example]]
name = "cookbook_structured"
path = "examples/cookbook/structured/main.rs"
[[example]]
name = "cookbook_multiturn"
path = "examples/cookbook/multiturn/main.rs"
[[example]]
name = "cookbook_agent"
path = "examples/cookbook/agent/main.rs"
[package.metadata.docs.rs]
no-default-features = true
[features]
# NVIDIA CUDA GPU support
cuda = ["hanzo-engine/cuda"]
# Vulkan GPU support (AMD APU via hanzo-ml vulkan backend)
vulkan = ["hanzo-engine/vulkan"]
rocm = ["hanzo-engine/rocm"]
# cuDNN acceleration (requires cuda)
cudnn = ["hanzo-engine/cudnn"]
# Apple Metal GPU support (macOS/iOS)
metal = ["hanzo-engine/metal"]
# Flash Attention 2 kernels (requires cuda)
flash-attn = ["cuda", "hanzo-engine/flash-attn"]
# Apple Accelerate framework for BLAS operations
accelerate = ["hanzo-engine/accelerate"]
# Intel MKL acceleration for BLAS operations
mkl = ["hanzo-engine/mkl"]
# Multi-GPU support via NCCL (requires cuda)
nccl = ["hanzo-engine/nccl"]
# Use ring instead of aws-lc-rs for TLS
ring = ["hanzo-engine/ring"]
# Python code execution tool
code-execution = ["hanzo-engine/code-execution"]
[[example]]
name = "deepseek_v4_gguf"
path = "examples/deepseek_v4_gguf/main.rs"
[[example]]
name = "v4_cache_dump"
path = "examples/v4_cache_dump/main.rs"