| Unified conformance test library (#434)
* Extract fusor conformance library
* Include mismatch position in conformance errors
* Remove remaining crate-local fusor tests
* Add missing conformance coverage
* Broaden conformance test coverage
* Use variable-size conformance fuzz shapes
* fix formatting
* fix tests
* larger fuzz ranges
* more tests
* better cpu parity
* fix clippy
* fix formatting
* fix clippy
* refactor: replace clippy #[allow] suppressions with real fixes
Bundle args into structs (FlushBatch, MmaParams, TileALoadCtx, TileBLoadCtx,
AttentionInputs, BertShape, QMatMulFuzz), introduce CompareFut type alias for
the conformance comparator return type, and rewrite needless_range_loop sites
in tests/common/mod.rs to use iter_mut().enumerate().
* chore: ignore .claude scheduler artifacts
* fix(conformance): skip f16 tests on GPU adapters without SHADER_F16
Linux CI's lavapipe adapter doesn't expose wgpu::Features::SHADER_F16, so
the f16 shader fails to validate. Filter the device list per test rather
than removing GPU coverage entirely — Mac Metal still runs the f16 path.
* fix(conformance): bump inverse-trig tolerance to 1e-4 for lavapipe parity
Linux CI's lavapipe adapter diverges from libm by ~6e-5 on asin/acos/atanh/acosh
near their domain edges. 1e-4 covers the observed gap without masking real
regressions; the macOS Metal adapter still passes comfortably.
* fix(conformance): bump inverse-trig tolerance to 1e-3 for lavapipe drift
First CI run showed asin diverging 5.6e-5; second run hit 2.1e-4 on a
different fuzz seed. lavapipe asin precision is limited near the asymptotes
where the derivative blows up. 1e-3 covers observed drift; algorithmic
regressions would diverge by orders of magnitude.
* fix transcribe example
* fix(conformance): stabilize CI edge cases
* fix(ci): avoid brittle benchmark formatter
* fix(conformance): cover Windows WARP tanh drift
* fix tanh
* more ci fixes
* more software gpu backend fixes
* looser bounds for trig
* relative tolerance
* more relative comparisons
* passing on warp | 28 天前 |
| Unified conformance test library (#434)
* Extract fusor conformance library
* Include mismatch position in conformance errors
* Remove remaining crate-local fusor tests
* Add missing conformance coverage
* Broaden conformance test coverage
* Use variable-size conformance fuzz shapes
* fix formatting
* fix tests
* larger fuzz ranges
* more tests
* better cpu parity
* fix clippy
* fix formatting
* fix clippy
* refactor: replace clippy #[allow] suppressions with real fixes
Bundle args into structs (FlushBatch, MmaParams, TileALoadCtx, TileBLoadCtx,
AttentionInputs, BertShape, QMatMulFuzz), introduce CompareFut type alias for
the conformance comparator return type, and rewrite needless_range_loop sites
in tests/common/mod.rs to use iter_mut().enumerate().
* chore: ignore .claude scheduler artifacts
* fix(conformance): skip f16 tests on GPU adapters without SHADER_F16
Linux CI's lavapipe adapter doesn't expose wgpu::Features::SHADER_F16, so
the f16 shader fails to validate. Filter the device list per test rather
than removing GPU coverage entirely — Mac Metal still runs the f16 path.
* fix(conformance): bump inverse-trig tolerance to 1e-4 for lavapipe parity
Linux CI's lavapipe adapter diverges from libm by ~6e-5 on asin/acos/atanh/acosh
near their domain edges. 1e-4 covers the observed gap without masking real
regressions; the macOS Metal adapter still passes comfortably.
* fix(conformance): bump inverse-trig tolerance to 1e-3 for lavapipe drift
First CI run showed asin diverging 5.6e-5; second run hit 2.1e-4 on a
different fuzz seed. lavapipe asin precision is limited near the asymptotes
where the derivative blows up. 1e-3 covers observed drift; algorithmic
regressions would diverge by orders of magnitude.
* fix transcribe example
* fix(conformance): stabilize CI edge cases
* fix(ci): avoid brittle benchmark formatter
* fix(conformance): cover Windows WARP tanh drift
* fix tanh
* more ci fixes
* more software gpu backend fixes
* looser bounds for trig
* relative tolerance
* more relative comparisons
* passing on warp | 28 天前 |
| Unified conformance test library (#434)
* Extract fusor conformance library
* Include mismatch position in conformance errors
* Remove remaining crate-local fusor tests
* Add missing conformance coverage
* Broaden conformance test coverage
* Use variable-size conformance fuzz shapes
* fix formatting
* fix tests
* larger fuzz ranges
* more tests
* better cpu parity
* fix clippy
* fix formatting
* fix clippy
* refactor: replace clippy #[allow] suppressions with real fixes
Bundle args into structs (FlushBatch, MmaParams, TileALoadCtx, TileBLoadCtx,
AttentionInputs, BertShape, QMatMulFuzz), introduce CompareFut type alias for
the conformance comparator return type, and rewrite needless_range_loop sites
in tests/common/mod.rs to use iter_mut().enumerate().
* chore: ignore .claude scheduler artifacts
* fix(conformance): skip f16 tests on GPU adapters without SHADER_F16
Linux CI's lavapipe adapter doesn't expose wgpu::Features::SHADER_F16, so
the f16 shader fails to validate. Filter the device list per test rather
than removing GPU coverage entirely — Mac Metal still runs the f16 path.
* fix(conformance): bump inverse-trig tolerance to 1e-4 for lavapipe parity
Linux CI's lavapipe adapter diverges from libm by ~6e-5 on asin/acos/atanh/acosh
near their domain edges. 1e-4 covers the observed gap without masking real
regressions; the macOS Metal adapter still passes comfortably.
* fix(conformance): bump inverse-trig tolerance to 1e-3 for lavapipe drift
First CI run showed asin diverging 5.6e-5; second run hit 2.1e-4 on a
different fuzz seed. lavapipe asin precision is limited near the asymptotes
where the derivative blows up. 1e-3 covers observed drift; algorithmic
regressions would diverge by orders of magnitude.
* fix transcribe example
* fix(conformance): stabilize CI edge cases
* fix(ci): avoid brittle benchmark formatter
* fix(conformance): cover Windows WARP tanh drift
* fix tanh
* more ci fixes
* more software gpu backend fixes
* looser bounds for trig
* relative tolerance
* more relative comparisons
* passing on warp | 28 天前 |
| Unified conformance test library (#434)
* Extract fusor conformance library
* Include mismatch position in conformance errors
* Remove remaining crate-local fusor tests
* Add missing conformance coverage
* Broaden conformance test coverage
* Use variable-size conformance fuzz shapes
* fix formatting
* fix tests
* larger fuzz ranges
* more tests
* better cpu parity
* fix clippy
* fix formatting
* fix clippy
* refactor: replace clippy #[allow] suppressions with real fixes
Bundle args into structs (FlushBatch, MmaParams, TileALoadCtx, TileBLoadCtx,
AttentionInputs, BertShape, QMatMulFuzz), introduce CompareFut type alias for
the conformance comparator return type, and rewrite needless_range_loop sites
in tests/common/mod.rs to use iter_mut().enumerate().
* chore: ignore .claude scheduler artifacts
* fix(conformance): skip f16 tests on GPU adapters without SHADER_F16
Linux CI's lavapipe adapter doesn't expose wgpu::Features::SHADER_F16, so
the f16 shader fails to validate. Filter the device list per test rather
than removing GPU coverage entirely — Mac Metal still runs the f16 path.
* fix(conformance): bump inverse-trig tolerance to 1e-4 for lavapipe parity
Linux CI's lavapipe adapter diverges from libm by ~6e-5 on asin/acos/atanh/acosh
near their domain edges. 1e-4 covers the observed gap without masking real
regressions; the macOS Metal adapter still passes comfortably.
* fix(conformance): bump inverse-trig tolerance to 1e-3 for lavapipe drift
First CI run showed asin diverging 5.6e-5; second run hit 2.1e-4 on a
different fuzz seed. lavapipe asin precision is limited near the asymptotes
where the derivative blows up. 1e-3 covers observed drift; algorithmic
regressions would diverge by orders of magnitude.
* fix transcribe example
* fix(conformance): stabilize CI edge cases
* fix(ci): avoid brittle benchmark formatter
* fix(conformance): cover Windows WARP tanh drift
* fix tanh
* more ci fixes
* more software gpu backend fixes
* looser bounds for trig
* relative tolerance
* more relative comparisons
* passing on warp | 28 天前 |
| Unified conformance test library (#434)
* Extract fusor conformance library
* Include mismatch position in conformance errors
* Remove remaining crate-local fusor tests
* Add missing conformance coverage
* Broaden conformance test coverage
* Use variable-size conformance fuzz shapes
* fix formatting
* fix tests
* larger fuzz ranges
* more tests
* better cpu parity
* fix clippy
* fix formatting
* fix clippy
* refactor: replace clippy #[allow] suppressions with real fixes
Bundle args into structs (FlushBatch, MmaParams, TileALoadCtx, TileBLoadCtx,
AttentionInputs, BertShape, QMatMulFuzz), introduce CompareFut type alias for
the conformance comparator return type, and rewrite needless_range_loop sites
in tests/common/mod.rs to use iter_mut().enumerate().
* chore: ignore .claude scheduler artifacts
* fix(conformance): skip f16 tests on GPU adapters without SHADER_F16
Linux CI's lavapipe adapter doesn't expose wgpu::Features::SHADER_F16, so
the f16 shader fails to validate. Filter the device list per test rather
than removing GPU coverage entirely — Mac Metal still runs the f16 path.
* fix(conformance): bump inverse-trig tolerance to 1e-4 for lavapipe parity
Linux CI's lavapipe adapter diverges from libm by ~6e-5 on asin/acos/atanh/acosh
near their domain edges. 1e-4 covers the observed gap without masking real
regressions; the macOS Metal adapter still passes comfortably.
* fix(conformance): bump inverse-trig tolerance to 1e-3 for lavapipe drift
First CI run showed asin diverging 5.6e-5; second run hit 2.1e-4 on a
different fuzz seed. lavapipe asin precision is limited near the asymptotes
where the derivative blows up. 1e-3 covers observed drift; algorithmic
regressions would diverge by orders of magnitude.
* fix transcribe example
* fix(conformance): stabilize CI edge cases
* fix(ci): avoid brittle benchmark formatter
* fix(conformance): cover Windows WARP tanh drift
* fix tanh
* more ci fixes
* more software gpu backend fixes
* looser bounds for trig
* relative tolerance
* more relative comparisons
* passing on warp | 28 天前 |
| Unified conformance test library (#434)
* Extract fusor conformance library
* Include mismatch position in conformance errors
* Remove remaining crate-local fusor tests
* Add missing conformance coverage
* Broaden conformance test coverage
* Use variable-size conformance fuzz shapes
* fix formatting
* fix tests
* larger fuzz ranges
* more tests
* better cpu parity
* fix clippy
* fix formatting
* fix clippy
* refactor: replace clippy #[allow] suppressions with real fixes
Bundle args into structs (FlushBatch, MmaParams, TileALoadCtx, TileBLoadCtx,
AttentionInputs, BertShape, QMatMulFuzz), introduce CompareFut type alias for
the conformance comparator return type, and rewrite needless_range_loop sites
in tests/common/mod.rs to use iter_mut().enumerate().
* chore: ignore .claude scheduler artifacts
* fix(conformance): skip f16 tests on GPU adapters without SHADER_F16
Linux CI's lavapipe adapter doesn't expose wgpu::Features::SHADER_F16, so
the f16 shader fails to validate. Filter the device list per test rather
than removing GPU coverage entirely — Mac Metal still runs the f16 path.
* fix(conformance): bump inverse-trig tolerance to 1e-4 for lavapipe parity
Linux CI's lavapipe adapter diverges from libm by ~6e-5 on asin/acos/atanh/acosh
near their domain edges. 1e-4 covers the observed gap without masking real
regressions; the macOS Metal adapter still passes comfortably.
* fix(conformance): bump inverse-trig tolerance to 1e-3 for lavapipe drift
First CI run showed asin diverging 5.6e-5; second run hit 2.1e-4 on a
different fuzz seed. lavapipe asin precision is limited near the asymptotes
where the derivative blows up. 1e-3 covers observed drift; algorithmic
regressions would diverge by orders of magnitude.
* fix transcribe example
* fix(conformance): stabilize CI edge cases
* fix(ci): avoid brittle benchmark formatter
* fix(conformance): cover Windows WARP tanh drift
* fix tanh
* more ci fixes
* more software gpu backend fixes
* looser bounds for trig
* relative tolerance
* more relative comparisons
* passing on warp | 28 天前 |
| Unified conformance test library (#434)
* Extract fusor conformance library
* Include mismatch position in conformance errors
* Remove remaining crate-local fusor tests
* Add missing conformance coverage
* Broaden conformance test coverage
* Use variable-size conformance fuzz shapes
* fix formatting
* fix tests
* larger fuzz ranges
* more tests
* better cpu parity
* fix clippy
* fix formatting
* fix clippy
* refactor: replace clippy #[allow] suppressions with real fixes
Bundle args into structs (FlushBatch, MmaParams, TileALoadCtx, TileBLoadCtx,
AttentionInputs, BertShape, QMatMulFuzz), introduce CompareFut type alias for
the conformance comparator return type, and rewrite needless_range_loop sites
in tests/common/mod.rs to use iter_mut().enumerate().
* chore: ignore .claude scheduler artifacts
* fix(conformance): skip f16 tests on GPU adapters without SHADER_F16
Linux CI's lavapipe adapter doesn't expose wgpu::Features::SHADER_F16, so
the f16 shader fails to validate. Filter the device list per test rather
than removing GPU coverage entirely — Mac Metal still runs the f16 path.
* fix(conformance): bump inverse-trig tolerance to 1e-4 for lavapipe parity
Linux CI's lavapipe adapter diverges from libm by ~6e-5 on asin/acos/atanh/acosh
near their domain edges. 1e-4 covers the observed gap without masking real
regressions; the macOS Metal adapter still passes comfortably.
* fix(conformance): bump inverse-trig tolerance to 1e-3 for lavapipe drift
First CI run showed asin diverging 5.6e-5; second run hit 2.1e-4 on a
different fuzz seed. lavapipe asin precision is limited near the asymptotes
where the derivative blows up. 1e-3 covers observed drift; algorithmic
regressions would diverge by orders of magnitude.
* fix transcribe example
* fix(conformance): stabilize CI edge cases
* fix(ci): avoid brittle benchmark formatter
* fix(conformance): cover Windows WARP tanh drift
* fix tanh
* more ci fixes
* more software gpu backend fixes
* looser bounds for trig
* relative tolerance
* more relative comparisons
* passing on warp | 28 天前 |
| Unified conformance test library (#434)
* Extract fusor conformance library
* Include mismatch position in conformance errors
* Remove remaining crate-local fusor tests
* Add missing conformance coverage
* Broaden conformance test coverage
* Use variable-size conformance fuzz shapes
* fix formatting
* fix tests
* larger fuzz ranges
* more tests
* better cpu parity
* fix clippy
* fix formatting
* fix clippy
* refactor: replace clippy #[allow] suppressions with real fixes
Bundle args into structs (FlushBatch, MmaParams, TileALoadCtx, TileBLoadCtx,
AttentionInputs, BertShape, QMatMulFuzz), introduce CompareFut type alias for
the conformance comparator return type, and rewrite needless_range_loop sites
in tests/common/mod.rs to use iter_mut().enumerate().
* chore: ignore .claude scheduler artifacts
* fix(conformance): skip f16 tests on GPU adapters without SHADER_F16
Linux CI's lavapipe adapter doesn't expose wgpu::Features::SHADER_F16, so
the f16 shader fails to validate. Filter the device list per test rather
than removing GPU coverage entirely — Mac Metal still runs the f16 path.
* fix(conformance): bump inverse-trig tolerance to 1e-4 for lavapipe parity
Linux CI's lavapipe adapter diverges from libm by ~6e-5 on asin/acos/atanh/acosh
near their domain edges. 1e-4 covers the observed gap without masking real
regressions; the macOS Metal adapter still passes comfortably.
* fix(conformance): bump inverse-trig tolerance to 1e-3 for lavapipe drift
First CI run showed asin diverging 5.6e-5; second run hit 2.1e-4 on a
different fuzz seed. lavapipe asin precision is limited near the asymptotes
where the derivative blows up. 1e-3 covers observed drift; algorithmic
regressions would diverge by orders of magnitude.
* fix transcribe example
* fix(conformance): stabilize CI edge cases
* fix(ci): avoid brittle benchmark formatter
* fix(conformance): cover Windows WARP tanh drift
* fix tanh
* more ci fixes
* more software gpu backend fixes
* looser bounds for trig
* relative tolerance
* more relative comparisons
* passing on warp | 28 天前 |
| Floneum cpu (#424)
* fix the summary chunker
* start
* wip
* some progress
* abox
* start simding
* add some operations
* pull out ResolveTensor
* refactor
* more refactoring
* fill in missing ops
* remove ResolvedTensorMut
* add benchmarks
* faster
* matmul op
* reduce operations
* refactor into multiple files
* fuse ops
* more ops
* clean up comments
* remove unsafe
* remove more unsafe
* initial fusor crate
* refactoring fusor cpu
* tensor type
* move more ops to tensor
* more refactoring
* more refactoring
* clean up some unused code
* remove most of fusor
* slice assign op
* quantized cpu
* optimize qmatmul
* wip
* new types crate
* use layout type in cpu
* pull out rank
* as_slice for gpuor
* better add impl
* partially borrowed ops
* refactor pairwise
* more gpuor ops
* more ops
* batched matmul
* implement some composite ops
* reduce ops
* normalization ops
* reshape ops
* move most of the logic onto layout itself
* move sliding window into the layout type
* a bunch more composite ops
* more methods
* more ops
* rename qmatmul
* batched cpu qmatmul
* quantized support for gpuor
* rename to be closer to fusor-core
* start porting rwhisper to cpu
* wip whisper
* optimize a bit
* more optimization
* more optimization
* rename eval
* more lazy
* remove some logs
* move all matches to dispatch
* more dispatch
* fix add_ and other ops
* use layout instead
* get rid of ResolveTensor
* remove useless Expr methods
* fix fusor
* shape on cpu tensor
* as ref tensor
* remove Expr trait
* remove expr from fusor
* owned and copy cpu tensors
* fix fusor/fusor
* start migrating fusor
* fix fusor
* rwhisper compiles
* fix reshape
* fix cpu map layout
* remove debug
* remove submodule
* simd qmatmul
* nightly optimizations
* switch llama to fusor
* remove const generic from QMatrix
* don't use rayon for parallelization
* MapLayout is currently a concrete tensor
* simd gather
* less matching in slow softmax
* wip
* remove SimdComparisonOp
* fix fusor
* lazy maplayout
* fix maplayout
* rwhisper compiles
* fix rwhisper
* wip
* llama running!
* fix vision on cpu
* wip q5k
* start gpu impl
* fix out of memory
* q5k_sgemv
* fix formatting
* fix some warnings
* remove some dead code
* refactor from array
* Use rustversion to conditionally enable nightly NEON intrinsics
The vdotq_s32 intrinsic requires the nightly-only
stdarch_neon_dotprod feature. This change uses rustversion
to detect nightly builds and only enables the optimized
NEON code path when building with nightly Rust. On stable
Rust, the scalar fallback implementation is used instead.
* fix formatting
* fix nightly formatting
* import SimdElement trait for x86 gather operations
* fix unsafe blocks for Rust 2024 edition
* fix clippy warnings in test modules
* fix tensor shape handling in conv1d and layer_norm loading
* skip qwen download in CI
* fix gelu in warp | 3 个月前 |
| Unified conformance test library (#434)
* Extract fusor conformance library
* Include mismatch position in conformance errors
* Remove remaining crate-local fusor tests
* Add missing conformance coverage
* Broaden conformance test coverage
* Use variable-size conformance fuzz shapes
* fix formatting
* fix tests
* larger fuzz ranges
* more tests
* better cpu parity
* fix clippy
* fix formatting
* fix clippy
* refactor: replace clippy #[allow] suppressions with real fixes
Bundle args into structs (FlushBatch, MmaParams, TileALoadCtx, TileBLoadCtx,
AttentionInputs, BertShape, QMatMulFuzz), introduce CompareFut type alias for
the conformance comparator return type, and rewrite needless_range_loop sites
in tests/common/mod.rs to use iter_mut().enumerate().
* chore: ignore .claude scheduler artifacts
* fix(conformance): skip f16 tests on GPU adapters without SHADER_F16
Linux CI's lavapipe adapter doesn't expose wgpu::Features::SHADER_F16, so
the f16 shader fails to validate. Filter the device list per test rather
than removing GPU coverage entirely — Mac Metal still runs the f16 path.
* fix(conformance): bump inverse-trig tolerance to 1e-4 for lavapipe parity
Linux CI's lavapipe adapter diverges from libm by ~6e-5 on asin/acos/atanh/acosh
near their domain edges. 1e-4 covers the observed gap without masking real
regressions; the macOS Metal adapter still passes comfortably.
* fix(conformance): bump inverse-trig tolerance to 1e-3 for lavapipe drift
First CI run showed asin diverging 5.6e-5; second run hit 2.1e-4 on a
different fuzz seed. lavapipe asin precision is limited near the asymptotes
where the derivative blows up. 1e-3 covers observed drift; algorithmic
regressions would diverge by orders of magnitude.
* fix transcribe example
* fix(conformance): stabilize CI edge cases
* fix(ci): avoid brittle benchmark formatter
* fix(conformance): cover Windows WARP tanh drift
* fix tanh
* more ci fixes
* more software gpu backend fixes
* looser bounds for trig
* relative tolerance
* more relative comparisons
* passing on warp | 28 天前 |
| Unified conformance test library (#434)
* Extract fusor conformance library
* Include mismatch position in conformance errors
* Remove remaining crate-local fusor tests
* Add missing conformance coverage
* Broaden conformance test coverage
* Use variable-size conformance fuzz shapes
* fix formatting
* fix tests
* larger fuzz ranges
* more tests
* better cpu parity
* fix clippy
* fix formatting
* fix clippy
* refactor: replace clippy #[allow] suppressions with real fixes
Bundle args into structs (FlushBatch, MmaParams, TileALoadCtx, TileBLoadCtx,
AttentionInputs, BertShape, QMatMulFuzz), introduce CompareFut type alias for
the conformance comparator return type, and rewrite needless_range_loop sites
in tests/common/mod.rs to use iter_mut().enumerate().
* chore: ignore .claude scheduler artifacts
* fix(conformance): skip f16 tests on GPU adapters without SHADER_F16
Linux CI's lavapipe adapter doesn't expose wgpu::Features::SHADER_F16, so
the f16 shader fails to validate. Filter the device list per test rather
than removing GPU coverage entirely — Mac Metal still runs the f16 path.
* fix(conformance): bump inverse-trig tolerance to 1e-4 for lavapipe parity
Linux CI's lavapipe adapter diverges from libm by ~6e-5 on asin/acos/atanh/acosh
near their domain edges. 1e-4 covers the observed gap without masking real
regressions; the macOS Metal adapter still passes comfortably.
* fix(conformance): bump inverse-trig tolerance to 1e-3 for lavapipe drift
First CI run showed asin diverging 5.6e-5; second run hit 2.1e-4 on a
different fuzz seed. lavapipe asin precision is limited near the asymptotes
where the derivative blows up. 1e-3 covers observed drift; algorithmic
regressions would diverge by orders of magnitude.
* fix transcribe example
* fix(conformance): stabilize CI edge cases
* fix(ci): avoid brittle benchmark formatter
* fix(conformance): cover Windows WARP tanh drift
* fix tanh
* more ci fixes
* more software gpu backend fixes
* looser bounds for trig
* relative tolerance
* more relative comparisons
* passing on warp | 28 天前 |
| Unified conformance test library (#434)
* Extract fusor conformance library
* Include mismatch position in conformance errors
* Remove remaining crate-local fusor tests
* Add missing conformance coverage
* Broaden conformance test coverage
* Use variable-size conformance fuzz shapes
* fix formatting
* fix tests
* larger fuzz ranges
* more tests
* better cpu parity
* fix clippy
* fix formatting
* fix clippy
* refactor: replace clippy #[allow] suppressions with real fixes
Bundle args into structs (FlushBatch, MmaParams, TileALoadCtx, TileBLoadCtx,
AttentionInputs, BertShape, QMatMulFuzz), introduce CompareFut type alias for
the conformance comparator return type, and rewrite needless_range_loop sites
in tests/common/mod.rs to use iter_mut().enumerate().
* chore: ignore .claude scheduler artifacts
* fix(conformance): skip f16 tests on GPU adapters without SHADER_F16
Linux CI's lavapipe adapter doesn't expose wgpu::Features::SHADER_F16, so
the f16 shader fails to validate. Filter the device list per test rather
than removing GPU coverage entirely — Mac Metal still runs the f16 path.
* fix(conformance): bump inverse-trig tolerance to 1e-4 for lavapipe parity
Linux CI's lavapipe adapter diverges from libm by ~6e-5 on asin/acos/atanh/acosh
near their domain edges. 1e-4 covers the observed gap without masking real
regressions; the macOS Metal adapter still passes comfortably.
* fix(conformance): bump inverse-trig tolerance to 1e-3 for lavapipe drift
First CI run showed asin diverging 5.6e-5; second run hit 2.1e-4 on a
different fuzz seed. lavapipe asin precision is limited near the asymptotes
where the derivative blows up. 1e-3 covers observed drift; algorithmic
regressions would diverge by orders of magnitude.
* fix transcribe example
* fix(conformance): stabilize CI edge cases
* fix(ci): avoid brittle benchmark formatter
* fix(conformance): cover Windows WARP tanh drift
* fix tanh
* more ci fixes
* more software gpu backend fixes
* looser bounds for trig
* relative tolerance
* more relative comparisons
* passing on warp | 28 天前 |
| Unified conformance test library (#434)
* Extract fusor conformance library
* Include mismatch position in conformance errors
* Remove remaining crate-local fusor tests
* Add missing conformance coverage
* Broaden conformance test coverage
* Use variable-size conformance fuzz shapes
* fix formatting
* fix tests
* larger fuzz ranges
* more tests
* better cpu parity
* fix clippy
* fix formatting
* fix clippy
* refactor: replace clippy #[allow] suppressions with real fixes
Bundle args into structs (FlushBatch, MmaParams, TileALoadCtx, TileBLoadCtx,
AttentionInputs, BertShape, QMatMulFuzz), introduce CompareFut type alias for
the conformance comparator return type, and rewrite needless_range_loop sites
in tests/common/mod.rs to use iter_mut().enumerate().
* chore: ignore .claude scheduler artifacts
* fix(conformance): skip f16 tests on GPU adapters without SHADER_F16
Linux CI's lavapipe adapter doesn't expose wgpu::Features::SHADER_F16, so
the f16 shader fails to validate. Filter the device list per test rather
than removing GPU coverage entirely — Mac Metal still runs the f16 path.
* fix(conformance): bump inverse-trig tolerance to 1e-4 for lavapipe parity
Linux CI's lavapipe adapter diverges from libm by ~6e-5 on asin/acos/atanh/acosh
near their domain edges. 1e-4 covers the observed gap without masking real
regressions; the macOS Metal adapter still passes comfortably.
* fix(conformance): bump inverse-trig tolerance to 1e-3 for lavapipe drift
First CI run showed asin diverging 5.6e-5; second run hit 2.1e-4 on a
different fuzz seed. lavapipe asin precision is limited near the asymptotes
where the derivative blows up. 1e-3 covers observed drift; algorithmic
regressions would diverge by orders of magnitude.
* fix transcribe example
* fix(conformance): stabilize CI edge cases
* fix(ci): avoid brittle benchmark formatter
* fix(conformance): cover Windows WARP tanh drift
* fix tanh
* more ci fixes
* more software gpu backend fixes
* looser bounds for trig
* relative tolerance
* more relative comparisons
* passing on warp | 28 天前 |
| Floneum cpu (#424)
* fix the summary chunker
* start
* wip
* some progress
* abox
* start simding
* add some operations
* pull out ResolveTensor
* refactor
* more refactoring
* fill in missing ops
* remove ResolvedTensorMut
* add benchmarks
* faster
* matmul op
* reduce operations
* refactor into multiple files
* fuse ops
* more ops
* clean up comments
* remove unsafe
* remove more unsafe
* initial fusor crate
* refactoring fusor cpu
* tensor type
* move more ops to tensor
* more refactoring
* more refactoring
* clean up some unused code
* remove most of fusor
* slice assign op
* quantized cpu
* optimize qmatmul
* wip
* new types crate
* use layout type in cpu
* pull out rank
* as_slice for gpuor
* better add impl
* partially borrowed ops
* refactor pairwise
* more gpuor ops
* more ops
* batched matmul
* implement some composite ops
* reduce ops
* normalization ops
* reshape ops
* move most of the logic onto layout itself
* move sliding window into the layout type
* a bunch more composite ops
* more methods
* more ops
* rename qmatmul
* batched cpu qmatmul
* quantized support for gpuor
* rename to be closer to fusor-core
* start porting rwhisper to cpu
* wip whisper
* optimize a bit
* more optimization
* more optimization
* rename eval
* more lazy
* remove some logs
* move all matches to dispatch
* more dispatch
* fix add_ and other ops
* use layout instead
* get rid of ResolveTensor
* remove useless Expr methods
* fix fusor
* shape on cpu tensor
* as ref tensor
* remove Expr trait
* remove expr from fusor
* owned and copy cpu tensors
* fix fusor/fusor
* start migrating fusor
* fix fusor
* rwhisper compiles
* fix reshape
* fix cpu map layout
* remove debug
* remove submodule
* simd qmatmul
* nightly optimizations
* switch llama to fusor
* remove const generic from QMatrix
* don't use rayon for parallelization
* MapLayout is currently a concrete tensor
* simd gather
* less matching in slow softmax
* wip
* remove SimdComparisonOp
* fix fusor
* lazy maplayout
* fix maplayout
* rwhisper compiles
* fix rwhisper
* wip
* llama running!
* fix vision on cpu
* wip q5k
* start gpu impl
* fix out of memory
* q5k_sgemv
* fix formatting
* fix some warnings
* remove some dead code
* refactor from array
* Use rustversion to conditionally enable nightly NEON intrinsics
The vdotq_s32 intrinsic requires the nightly-only
stdarch_neon_dotprod feature. This change uses rustversion
to detect nightly builds and only enables the optimized
NEON code path when building with nightly Rust. On stable
Rust, the scalar fallback implementation is used instead.
* fix formatting
* fix nightly formatting
* import SimdElement trait for x86 gather operations
* fix unsafe blocks for Rust 2024 edition
* fix clippy warnings in test modules
* fix tensor shape handling in conv1d and layer_norm loading
* skip qwen download in CI
* fix gelu in warp | 3 个月前 |
| Floneum cpu (#424)
* fix the summary chunker
* start
* wip
* some progress
* abox
* start simding
* add some operations
* pull out ResolveTensor
* refactor
* more refactoring
* fill in missing ops
* remove ResolvedTensorMut
* add benchmarks
* faster
* matmul op
* reduce operations
* refactor into multiple files
* fuse ops
* more ops
* clean up comments
* remove unsafe
* remove more unsafe
* initial fusor crate
* refactoring fusor cpu
* tensor type
* move more ops to tensor
* more refactoring
* more refactoring
* clean up some unused code
* remove most of fusor
* slice assign op
* quantized cpu
* optimize qmatmul
* wip
* new types crate
* use layout type in cpu
* pull out rank
* as_slice for gpuor
* better add impl
* partially borrowed ops
* refactor pairwise
* more gpuor ops
* more ops
* batched matmul
* implement some composite ops
* reduce ops
* normalization ops
* reshape ops
* move most of the logic onto layout itself
* move sliding window into the layout type
* a bunch more composite ops
* more methods
* more ops
* rename qmatmul
* batched cpu qmatmul
* quantized support for gpuor
* rename to be closer to fusor-core
* start porting rwhisper to cpu
* wip whisper
* optimize a bit
* more optimization
* more optimization
* rename eval
* more lazy
* remove some logs
* move all matches to dispatch
* more dispatch
* fix add_ and other ops
* use layout instead
* get rid of ResolveTensor
* remove useless Expr methods
* fix fusor
* shape on cpu tensor
* as ref tensor
* remove Expr trait
* remove expr from fusor
* owned and copy cpu tensors
* fix fusor/fusor
* start migrating fusor
* fix fusor
* rwhisper compiles
* fix reshape
* fix cpu map layout
* remove debug
* remove submodule
* simd qmatmul
* nightly optimizations
* switch llama to fusor
* remove const generic from QMatrix
* don't use rayon for parallelization
* MapLayout is currently a concrete tensor
* simd gather
* less matching in slow softmax
* wip
* remove SimdComparisonOp
* fix fusor
* lazy maplayout
* fix maplayout
* rwhisper compiles
* fix rwhisper
* wip
* llama running!
* fix vision on cpu
* wip q5k
* start gpu impl
* fix out of memory
* q5k_sgemv
* fix formatting
* fix some warnings
* remove some dead code
* refactor from array
* Use rustversion to conditionally enable nightly NEON intrinsics
The vdotq_s32 intrinsic requires the nightly-only
stdarch_neon_dotprod feature. This change uses rustversion
to detect nightly builds and only enables the optimized
NEON code path when building with nightly Rust. On stable
Rust, the scalar fallback implementation is used instead.
* fix formatting
* fix nightly formatting
* import SimdElement trait for x86 gather operations
* fix unsafe blocks for Rust 2024 edition
* fix clippy warnings in test modules
* fix tensor shape handling in conv1d and layer_norm loading
* skip qwen download in CI
* fix gelu in warp | 3 个月前 |
| Unified conformance test library (#434)
* Extract fusor conformance library
* Include mismatch position in conformance errors
* Remove remaining crate-local fusor tests
* Add missing conformance coverage
* Broaden conformance test coverage
* Use variable-size conformance fuzz shapes
* fix formatting
* fix tests
* larger fuzz ranges
* more tests
* better cpu parity
* fix clippy
* fix formatting
* fix clippy
* refactor: replace clippy #[allow] suppressions with real fixes
Bundle args into structs (FlushBatch, MmaParams, TileALoadCtx, TileBLoadCtx,
AttentionInputs, BertShape, QMatMulFuzz), introduce CompareFut type alias for
the conformance comparator return type, and rewrite needless_range_loop sites
in tests/common/mod.rs to use iter_mut().enumerate().
* chore: ignore .claude scheduler artifacts
* fix(conformance): skip f16 tests on GPU adapters without SHADER_F16
Linux CI's lavapipe adapter doesn't expose wgpu::Features::SHADER_F16, so
the f16 shader fails to validate. Filter the device list per test rather
than removing GPU coverage entirely — Mac Metal still runs the f16 path.
* fix(conformance): bump inverse-trig tolerance to 1e-4 for lavapipe parity
Linux CI's lavapipe adapter diverges from libm by ~6e-5 on asin/acos/atanh/acosh
near their domain edges. 1e-4 covers the observed gap without masking real
regressions; the macOS Metal adapter still passes comfortably.
* fix(conformance): bump inverse-trig tolerance to 1e-3 for lavapipe drift
First CI run showed asin diverging 5.6e-5; second run hit 2.1e-4 on a
different fuzz seed. lavapipe asin precision is limited near the asymptotes
where the derivative blows up. 1e-3 covers observed drift; algorithmic
regressions would diverge by orders of magnitude.
* fix transcribe example
* fix(conformance): stabilize CI edge cases
* fix(ci): avoid brittle benchmark formatter
* fix(conformance): cover Windows WARP tanh drift
* fix tanh
* more ci fixes
* more software gpu backend fixes
* looser bounds for trig
* relative tolerance
* more relative comparisons
* passing on warp | 28 天前 |
| Unified conformance test library (#434)
* Extract fusor conformance library
* Include mismatch position in conformance errors
* Remove remaining crate-local fusor tests
* Add missing conformance coverage
* Broaden conformance test coverage
* Use variable-size conformance fuzz shapes
* fix formatting
* fix tests
* larger fuzz ranges
* more tests
* better cpu parity
* fix clippy
* fix formatting
* fix clippy
* refactor: replace clippy #[allow] suppressions with real fixes
Bundle args into structs (FlushBatch, MmaParams, TileALoadCtx, TileBLoadCtx,
AttentionInputs, BertShape, QMatMulFuzz), introduce CompareFut type alias for
the conformance comparator return type, and rewrite needless_range_loop sites
in tests/common/mod.rs to use iter_mut().enumerate().
* chore: ignore .claude scheduler artifacts
* fix(conformance): skip f16 tests on GPU adapters without SHADER_F16
Linux CI's lavapipe adapter doesn't expose wgpu::Features::SHADER_F16, so
the f16 shader fails to validate. Filter the device list per test rather
than removing GPU coverage entirely — Mac Metal still runs the f16 path.
* fix(conformance): bump inverse-trig tolerance to 1e-4 for lavapipe parity
Linux CI's lavapipe adapter diverges from libm by ~6e-5 on asin/acos/atanh/acosh
near their domain edges. 1e-4 covers the observed gap without masking real
regressions; the macOS Metal adapter still passes comfortably.
* fix(conformance): bump inverse-trig tolerance to 1e-3 for lavapipe drift
First CI run showed asin diverging 5.6e-5; second run hit 2.1e-4 on a
different fuzz seed. lavapipe asin precision is limited near the asymptotes
where the derivative blows up. 1e-3 covers observed drift; algorithmic
regressions would diverge by orders of magnitude.
* fix transcribe example
* fix(conformance): stabilize CI edge cases
* fix(ci): avoid brittle benchmark formatter
* fix(conformance): cover Windows WARP tanh drift
* fix tanh
* more ci fixes
* more software gpu backend fixes
* looser bounds for trig
* relative tolerance
* more relative comparisons
* passing on warp | 28 天前 |
| Unified conformance test library (#434)
* Extract fusor conformance library
* Include mismatch position in conformance errors
* Remove remaining crate-local fusor tests
* Add missing conformance coverage
* Broaden conformance test coverage
* Use variable-size conformance fuzz shapes
* fix formatting
* fix tests
* larger fuzz ranges
* more tests
* better cpu parity
* fix clippy
* fix formatting
* fix clippy
* refactor: replace clippy #[allow] suppressions with real fixes
Bundle args into structs (FlushBatch, MmaParams, TileALoadCtx, TileBLoadCtx,
AttentionInputs, BertShape, QMatMulFuzz), introduce CompareFut type alias for
the conformance comparator return type, and rewrite needless_range_loop sites
in tests/common/mod.rs to use iter_mut().enumerate().
* chore: ignore .claude scheduler artifacts
* fix(conformance): skip f16 tests on GPU adapters without SHADER_F16
Linux CI's lavapipe adapter doesn't expose wgpu::Features::SHADER_F16, so
the f16 shader fails to validate. Filter the device list per test rather
than removing GPU coverage entirely — Mac Metal still runs the f16 path.
* fix(conformance): bump inverse-trig tolerance to 1e-4 for lavapipe parity
Linux CI's lavapipe adapter diverges from libm by ~6e-5 on asin/acos/atanh/acosh
near their domain edges. 1e-4 covers the observed gap without masking real
regressions; the macOS Metal adapter still passes comfortably.
* fix(conformance): bump inverse-trig tolerance to 1e-3 for lavapipe drift
First CI run showed asin diverging 5.6e-5; second run hit 2.1e-4 on a
different fuzz seed. lavapipe asin precision is limited near the asymptotes
where the derivative blows up. 1e-3 covers observed drift; algorithmic
regressions would diverge by orders of magnitude.
* fix transcribe example
* fix(conformance): stabilize CI edge cases
* fix(ci): avoid brittle benchmark formatter
* fix(conformance): cover Windows WARP tanh drift
* fix tanh
* more ci fixes
* more software gpu backend fixes
* looser bounds for trig
* relative tolerance
* more relative comparisons
* passing on warp | 28 天前 |
| Unified conformance test library (#434)
* Extract fusor conformance library
* Include mismatch position in conformance errors
* Remove remaining crate-local fusor tests
* Add missing conformance coverage
* Broaden conformance test coverage
* Use variable-size conformance fuzz shapes
* fix formatting
* fix tests
* larger fuzz ranges
* more tests
* better cpu parity
* fix clippy
* fix formatting
* fix clippy
* refactor: replace clippy #[allow] suppressions with real fixes
Bundle args into structs (FlushBatch, MmaParams, TileALoadCtx, TileBLoadCtx,
AttentionInputs, BertShape, QMatMulFuzz), introduce CompareFut type alias for
the conformance comparator return type, and rewrite needless_range_loop sites
in tests/common/mod.rs to use iter_mut().enumerate().
* chore: ignore .claude scheduler artifacts
* fix(conformance): skip f16 tests on GPU adapters without SHADER_F16
Linux CI's lavapipe adapter doesn't expose wgpu::Features::SHADER_F16, so
the f16 shader fails to validate. Filter the device list per test rather
than removing GPU coverage entirely — Mac Metal still runs the f16 path.
* fix(conformance): bump inverse-trig tolerance to 1e-4 for lavapipe parity
Linux CI's lavapipe adapter diverges from libm by ~6e-5 on asin/acos/atanh/acosh
near their domain edges. 1e-4 covers the observed gap without masking real
regressions; the macOS Metal adapter still passes comfortably.
* fix(conformance): bump inverse-trig tolerance to 1e-3 for lavapipe drift
First CI run showed asin diverging 5.6e-5; second run hit 2.1e-4 on a
different fuzz seed. lavapipe asin precision is limited near the asymptotes
where the derivative blows up. 1e-3 covers observed drift; algorithmic
regressions would diverge by orders of magnitude.
* fix transcribe example
* fix(conformance): stabilize CI edge cases
* fix(ci): avoid brittle benchmark formatter
* fix(conformance): cover Windows WARP tanh drift
* fix tanh
* more ci fixes
* more software gpu backend fixes
* looser bounds for trig
* relative tolerance
* more relative comparisons
* passing on warp | 28 天前 |
| Get whisper running in wasm and enable support for gpus without subgroups or f16 (#406)
* make rwhisper wasm compatable
* move around traits to minimize whisper dependencies
* fix transcription task
* fix import FutureWasmNotSend
* more wasm fixes
* fix sgemv
* reduce with or without subgroups
* fix subgroup reduction
* softmax without subgroups
* tiled map without subgroups
* quantized without subgroups
* all tests passing without subgroups
* fix sgemv dispatch
* restore subgroupless CI workflows
* only require f16 support for quantization support
* don't require f16 support for qmatmul
* fix kalosm model-types dev dependancies
* fix kalosm-model-types tests
* fix softmax bounds without subgroups
* better required limits
* fix clippy
* more clippy fixes
* shorter odyssey test
* more clippy fixes
* fix solve when workgroups are not supported
* exclude fusor core - these tests pass locally, but are too slow to run
in CI
* exclude inference tests on windows
* exclude kalosm-learning on windows which doesn't have f16 support | 6 个月前 |
| Floneum qwen embed and rbert cpu support (#425)
* qwen embed working
* Add 3×4 outer-loop unrolled quantized matmul for m≥2
Process 3 LHS rows simultaneously so each weight block is loaded
once and reused across all 3 rows, reducing memory traffic by ~3×.
2-4.4× speedup for prompt processing (m≥2); m=1 path unchanged.
* Deduplicate activation quantization in parallel m=1 path
Each thread now quantizes the activation row once instead of
re-quantizing in every 32-column chunk. 14-23% speedup for m=1.
* Add x86 AVX2 SIMD for BlockQ4_0 and BlockQ8_0 vec_dot
Use _mm256_maddubs_epi16 with the sign trick (abs(x) * sign(y,x))
for signed i8×i8 dot products. Runtime AVX2 detection with scalar
fallback for older x86_64 CPUs.
* Add x86 AVX2 SIMD for BlockQ8_0 activation quantization
Use AVX2 for the full quantize pipeline on x86_64: max-abs reduction,
scale+round via _mm256_cvtps_epi32, and saturating pack i32→i16→i8.
Runtime AVX2 detection with scalar fallback. On aarch64, the compiler
auto-vectorizes the scalar code better than explicit NEON intrinsics
(confirmed by benchmarks showing 7-11% regression with explicit NEON).
* Remove unused process_row_integer_range function
* cpu support for rbert
* better parallelization
* remove uninit unchecked
* start refactoring
* reduce unsafe
* fix formatting
* clean up conditional
* more refactoring
* a bit more cleanup
* more formatting + clippy
* fix clippy in fusion bench
* fix flash attention
* fix flash
* fix tests
* fix clippy
* make device week | 2 个月前 |
| Bump wgpu and fix wasm support for llama (#416)
* bump wgpu and fix wasm support for llama
* use git version of wgpu
* fix checks and require tokio for ocr
* fix formatting
* fix the quantized test
* fix large dispatches
* fix doctests
* fix clippy
* match wgpu version in ci and cache windows
* fix formatting
* setup vulkan
* set vulkan version
* cargo update
* failing tests on all platforms
* fix f16 tensors on gpus that don't support f16
* restore f16 detection
* more resilient caching
* pull out lock current logic
* use dxc
* use std file locks
* try using a smaller batch limit
* install warp | 4 个月前 |