Crate fuel_tensor 

fuel-tensor — Fuel’s user-facing Tensor handle and Device abstraction, moved here from fuel-core (board #109; CireSnave: “Move Tensor into fuel-tensor. We can worry about a semantic split later.”). This is a MECHANICAL move – every item below kept its own content; only the crate boundary changed. fuel-core’s own lib.rs re-exports these same paths so existing fuel_tensor::lazy::Tensor / fuel::Tensor call sites keep compiling.

The moved cluster (lazy, device, the 3 GPU backend bridges, dtype’s WithDType/IntDType/FloatDType trio, pipelined_bridge, judge, factories, planner, decode_shape, lazy_latent_cache, test_utils, scheduling, inference_context, kv_block_pool_device, persistent_decode, nf4) moved together in one PR because it is one mutually-referencing graph, not several independent files that happen to touch Tensor – measured via a full crate::-reference census before any file moved, not assumed from file names. See docs/release-0.13.0-wave.md.

use fuel_tensor::lazy::{Tensor, realize_many_f32};
use fuel_tensor::Device;
let dev = Device::cpu();
let a = Tensor::from_f32((0..6).map(|x| x as f32).collect::<Vec<_>>(), (2, 3), &dev)?;
let b = Tensor::from_f32_on(a.graph(), (0..12).map(|x| x as f32).collect::<Vec<_>>(), (3, 4), &dev)?;
let c = a.matmul(&b)?;
assert_eq!(c.shape().dims(), &[2, 4]);
let out = realize_many_f32(&[&c]);
assert_eq!(out[0].len(), 8);

Re-exports

pub use device::Device;

pub use device::NdArray;

pub use dtype::FloatDType;

pub use dtype::IntDType;

pub use dtype::WithDType;

Modules

awq

HuggingFace *-AWQ (GEMM-version) weight loading (ROADMAP row 31).

cuda_backend

Thin re-export layer for CUDA backend types, plus the bridge that turns a [fuel_cuda_backend::CudaDevice] into a crate::Device.

decode_shape

The identity a held decode plan is baked against.

device

Device abstraction for CPU, CUDA, and Metal backends.

dtype

Types for elements that can be stored and manipulated using tensors.

factories

Backend factory registry — Step 12 of the backend-agnostic refactor.

inference_context

Persistent inference state for the pipelined executor (Phase 7.6 step 9c, sub-phase E.3.0).

judge

The Judge — Phase 6b’s empirical profiler.

kv_block_pool_device

Device-backed KV block pool — the thin integration layer that binds the pure KvBlockPool host-side allocator to real device Storage (multi-session serving, Increment 2, Part 2).

lazy

Phase 6a bridge: a lazy-computation-graph tensor that wraps fuel_graph::NodeHandle and presents it through an API compatible with fuel-core’s eager Tensor.

lazy_latent_cache

LatentCache — a general N-slot decode cache backed by Tensor, the structural generalization of [fuel_transformers::models::lazy_kv_cache:: LazyKvCache] that unblocks latent / pruned KV compression.

metal_backend

Thin re-export layer for Metal backend types, plus the bridge that turns a [fuel_metal_backend::MetalDevice] into a crate::Device.

nf4

bitsandbytes NF4 weight materialization helpers.

persistent_decode

The shared persistent-decode rebind driver (GAP-029 increment 2b).

pipelined_bridge

Bridge from fuel-core’s user-facing Tensor::realize_* API to fuel-storage’s PipelinedExecutor (Phase 7.6 step 9c, sub-phases E.1 + E.2).

planner

Load-time planning driver.

reproducibility

Strict reproducibility mode (ROADMAP item 13(b)).

scheduling

Orchestrator for the Phase 6b probe → judge → dispatch pipeline.

test_utils

Test helpers. The eager-Tensor helpers (test_device!, assert_tensor_eq, to_vec{0,1,2,3}_round) were removed in B6; what remains is host-slice and lazy infrastructure that the lazy_* port tests use.

Macros

bail

Returns early from a function with a formatted error message.

Structs

Layout

Describes the memory layout of a tensor: shape, strides, and start offset.

Shape

A shape represents the dimensions of a tensor.

Storage

Owns a typed contiguous buffer on one device. The boxed DynBackendStorage is the actual byte holder; Storage is a thin wrapper that gives the eager-dispatch op methods (matmul, conv, unary, binary, …) somewhere to live.

Enums

D

Dimension indexing from the right (end) of a shape.

DType

The different types of elements allowed in tensors.

DTypeParseError

Error returned when a string cannot be parsed as a DType.

DeviceLocation

Identifies a physical device location (CPU or a specific GPU).

Error

The main error type for the fuel library.

HostBuffer

Host-addressable tensor storage holding a typed vector of elements.

Traits

Context

Attach more context to an error.

Type Aliases

Result

A specialized Result type for fuel operations.