0
Fork 0
mirror of https://github.com/zml/zml.git synced 2026-11-16 19:11:09 +00:00
Commit graph

1,093 commits

Author SHA1 Message Date
Erwan Vivien
5253ea1362 stdx/flags: cleanup flag_name using mem.replaceScalar 2026-11-13 11:36:07 +01:00
Erwan Vivien
bb2ded0e21 stdx/flags: add cheap tests 2026-11-13 11:29:26 +01:00
Erwan Vivien
7378ca8acf stdx/flags: support --flag value along --flag=value 2026-11-13 11:29:26 +01:00
Corentin Godeau
dc829cfb7d
[Modularization] zml/kernels/moe: replace options.zig with api.zig (#1059)
Also moved `LocalRouting` to `api.zig`
2026-11-13 10:50:15 +01:00
Kevin Barre
c40fa52d2d
zml/moe [Platform:tcl]: support block-128 FP8 SwiGLU experts (#1053)
- Select TCL for plain E4M3FN block-128 experts with matching BF16 or
  F32 scales. Preserve rejection of unsupported quantization.

- Quantize activations per block and apply checkpoint scales during
  native projections. Keep BF16 materialization and F32 down and
  routing accumulation.

- Reuse local expert routing and AllReduce for expert parallelism.
  Choose expert-stationary for small FP8 token counts and use the
  existing activation-stationary and blockwise recipes for larger
  counts.

- Share FP8 and BF16 dispatch preparation while retaining the token
  padding and BF16 SiLU rounding workarounds.
2026-11-12 13:47:48 +01:00
Guillaume Wenzek
8ebed54029
zml: refactor safetensors (#1048)
Refactor `zml/safetensors.zig`

* Drop metadata parsing
* refactor code to be more "object oriented", code is owned by the
`TensorRegistry`
* implement a new format `model.safetensors.full_index.json` that is
similar to `index.json` but actually has data inside it:

```
{
"metadata": {"num_tensors":96085,"total_size":510286023000},
"weight_map": {
  "aligner.w1.bias": {"filename":"model-00001-of-00048.safetensors","shape":[5120],"data_offsets":[27384,37624],"dtype":"BF16"},
  "aligner.w1.weight": {"filename":"model-00001-of-00048.safetensors","shape":[5120,9216],"data_offsets":[37624,94409464],"dtype":"BF16"},
  "aligner.w2.bias": {"filename":"model-00001-of-00048.safetensors","shape":[5120],"data_offsets":[94409464,94419704],"dtype":"BF16"},
  "aligner.w2.weight": {"filename":"model-00001-of-00048.safetensors","shape":[5120,5120],"data_offsets":[94419704,146848504],"dtype":"BF16"},
  "vision.blocks.0.attn.wo.bias": {"filename":"model-00001-of-00048.safetensors","shape":[1024],"data_offsets":[146848504,146850552],"dtype":"BF16"},
  "vision.blocks.0.attn.wo.weight": {"filename":"model-00001-of-00048.safetensors","shape":[1024,1024],"data_offsets":[146850552,148947704],"dtype":"BF16"},
  "vision.blocks.0.attn.wqkv.bias": {"filename":"model-00001-of-00048.safetensors","shape":[3072],"data_offsets":[148947704,148953848],"dtype":"BF16"},
 ...
 ```
 
 I followed conventions of `model.saftensors.index.json`
2026-11-12 13:12:25 +01:00
Raphael Nahoum
2b604ebfc6
zml/kernels/flash_attention/metal: port the kernels to the Metal DSL (#1054)
Replaces the `zml$flash_attn` custom call with Metal DSL kernels over
contiguous K/V:

- `fa_vec` + `fa_vec_reduce` for bf16 decode, at most 256 threads per
threadgroup.
- `fa_tiled` for prefill and f16.
- Batched queries work (the old path panicked). Other dtypes and head
dims fall back to sdpa.

Bumps the plugins to
[202611101651.363.1.d46c7471c419](2026111016.363.1.d46c7471c419)
2026-11-10 18:51:08 +01:00
Raphael Nahoum
9f8da93eaf
zml/kernels/paged_attention/metal: fit decode in 256 threads and split it over the GPU cores (#1046)
Fixes Metal decode on M1/M2: `fa_vec_paged` launched up to 512 threads
per threadgroup, over the 384 those GPUs accept.

- A threadgroup holds up to 4 query heads of one KV head times position
splits, at most 256 threads.
- The split count follows the PJRT core count and the number of rows.
- Head dims 32 and 64 now use the vector kernel instead of the tiled
one.

llmd decode on M5 Max vs master: no change at short context, +4–27% at
8K, +12–74% at 24K, bs16 unchanged.
Head-dim-64 models: 2–7.6× at long context.

Should fix https://github.com/zml/zml/issues/1037 but needs some testing
on a real M1/M2
2026-11-10 18:51:08 +01:00
Guillaume Wenzek
6ee4f84ea7
zml: Exe.memoryStats (#1052) 2026-11-10 14:24:41 +01:00
Louis Lechevalier
b61eede864
zml/kernels: Enable MegaMoe support in llmd (#1032)
This PR enable the MegaMoe kernel support in llmd by :
- Making the symmetric workspace an arg of CuteOps
- Adding cuda.grid_constant arg to the DSL builder, enabling GPU kernels
to receive TMA tensormap
- Adapting the existing mxfp4 kernel (non mega) to the new shared
swizzling `interleaved_8`
2026-11-10 14:23:27 +01:00
Raphael Nahoum
7b40b1b0f8
zml/io: honor slices and splat in DirectMemoryWriter.drain (#1050)
`DirectMemoryWriter.drain` only wrote the first slice of data and
ignored the splat count. The `std.Io.Writer` contract is to write every
slice in order, then repeat the last one splat times.

- A one-byte splat advanced one byte per drain call.
- After a partial write, `splatBytesAll` passes an empty first slice
followed by the pattern. The old drain returned 0 for that forever, so
the fill never finished.

`drain` now writes every slice and repeats the last one splat times.
2026-11-09 20:48:45 +01:00
Corentin Godeau
b23dbe334a
[Modularization] zml/kernels/moe: break cycle between flashinfer_cutlass and platform (#1042)
* Stopped using `zml.ops.CustomCall` because it creates cycles
impossible to break.
* renamed some `cutlass_flashinfer` to `flashinfer_cutlass` for
consistency

I didn't replace `CustomCall` because flashinfer_cutlass is the last
user and will be removed soon.
2026-11-09 19:24:38 +01:00
Corentin Godeau
2c57959202
[Modularization] zml/kernels/moe: break cycle (#1041)
`zml/kernels/moe.zig` is not imported by backend implementation anymore.

I introduced `options.zig` to contain the shared options. I also
introduced `weights.zig` to expose the unpacking helpers.
2026-11-09 19:24:38 +01:00
Corentin Godeau
b73de969eb
[Modularization] zml: break cycle in zml/kernels/paged_attention (#1039) 2026-11-09 19:24:37 +01:00
Corentin Godeau
7fbd15eee4
[Modularization] zml: break cycle in zml/attention (#1038)
Break cycle between `zml/attention/attention.zig` and
`zml/kernels/flash_attention/flash_attention.zig`
2026-11-09 19:24:37 +01:00
Hugo Mano
1042cba3ff
zml/kernels/moe: make mesh and partitioning explicit (#1028)
MOE API currently relies on input sharding metadata or
partitioning choices made inside backend kernels. Ordinary graph
operations can leave that metadata unspecified, making the resulting
sharding depend on the backend and preceding operations.

Move the manual computation boundary into the shared
`zml.kernels.moe.forwardMoe` API using `manualComputation2`. 
Callers provide a logical mesh and typed input/output partitioning specs.
2026-11-09 15:26:03 +01:00
Corentin Godeau
59a888f7e2
third_party/zls: allow all 0.17 zig versions (#1036) 2026-11-06 15:14:38 +01:00
Raphael Nahoum
5b079127bb
workspace: bump ZML plugins (#1034)
Bump plugins to
[202611060925.355.1.2ca211fa5fbd](2026110609.355.1.2ca211fa5fbd)
2026-11-06 10:56:05 +01:00
PierreLgol
488c21c899
workspace: add Helix editor configuration for Zig (#1029)
adding basic support for the helix editor, to automatically work
2026-11-06 10:29:31 +01:00
erwanvivien
837e45022a
zml/kernels/paged_attention: fix stablehlo backend slot_id increment (#1027)
If sequences were not 16-aligned we would have sequence leaks. Now we
increase slot_id by query_start_len and pad when needed
2026-11-06 09:57:09 +01:00
Corentin Godeau
630d96726b
zml/io: add consolidated version of safetensors to defaults (#1035)
Some models entrypoints are exposed as
`consolidated.safetensors.index.json` or `consolidated.safetensors`.
2026-11-05 19:34:17 +01:00
Raphael Nahoum
3285b7e0f4
zml: FP8 block activations and paged attention query padding on Furiosa and bump plugins (#1033)
- Quantize block FP8 activations to E4M3FN on Furiosa, so scaled dots
run natively. Other block sizes keep BF16 activations and XLA's
fallback.

- Pad paged attention query groups to a power of two with zero queries,
as TCC fails other group sizes. Marked as a workaround.

- Bump to
[202611051630.351.1.48d539621544](2026110516.351.1.48d539621544)
2026-11-05 17:38:31 +01:00
Corentin Godeau
5decef1fca
zml/kernels/paged_attention: enable cute by default when available (#1017) 2026-11-05 14:17:44 +01:00
Hugo Mano
2438b9c364
zml/kernels/paged_attention: make mesh and partitioning explicit (#1024)
Paged attention currently relies on input sharding metadata or
partitioning choices made inside backend kernels. Ordinary graph
operations can leave that metadata unspecified, making the resulting
sharding depend on the backend and preceding operations.

Move the manual computation boundary into the shared `pagedAttention`
API using `manualComputation2`. Callers provide a logical mesh and typed
input/output partitioning specs.
2026-11-05 13:19:58 +01:00
Hugo Mano
47f50667d8
zml/attention: declare manual partitioning in the shared API instead of kernels (#1016)
Move manual regions out of flash attention kernels and use
`manualComputation2` for every shared attention backend.

Also fully replicate the embedding executable outputs in Qwen and Llama.
Partial graph constraints leave their batch and sequence dimensions
open, which is invalid at executable boundaries.

Fix two further blockers exposed by the Qwen run that we forgot to move
recently: softplus comparisons must use the input dtype and ROCm DNN
runfiles must include the MIOpen kernel archives.
2026-11-05 13:00:00 +01:00
Guillaume Wenzek
4baec25924
kernels/moe: remove confusing bitcast (#1030)
the bitcast discard partitioning info
the API issue is tracked by https://github.com/zml/zml/issues/914
2026-11-05 11:53:31 +01:00
Louis Lechevalier
69a6b2bb22
workspace: bump ZML plugins (#1031) 2026-11-05 11:44:39 +01:00
Damien Mehala
8898051c7b
zml/kernels: fix Triton split-K (#1012)
The split-K/3D path writes partial results only for active queries, but
the reduction kernel processes inactive queries in the padded batch.
These queries have no valid results and cause invalid
computation/outputs.

This fix zeroes the ouput for inactive queries and return before before
accessing sequence metadata.

Resolves #1006
2026-11-04 20:48:56 +01:00
Raphael Nahoum
5764202e5f
zml/kernels/paged_attention/tcl: follow sequence lengths and support several chips (#1026)
- **Work follows sequence lengths:** decode's compute no longer runs
over the block table's whole width, and flash prefill loops only over
the chunks in use.

- **One plan per call:** each call picks straight decode, the hybrid
loop or flash attention, following a calibration sweep based on what
compiles. The kernels might change once the compiler evolves.

- **Several chips:** each chip runs the kernels over its own heads
inside a manual computation.
2026-11-04 18:52:05 +01:00
Guillaume Wenzek
db119b2d35
zml.Shape: improve parseTags (#1023)
avoid `stdx.BoundedArray` for tags, use `[]const Tag`, this allows to
simplify some logic and more explicitly separate comptime work from
runtime work.
It reduces binary sizes, preventing big functions to be duplicated
because of different tags.
2026-11-04 16:04:19 +01:00
Guillaume Wenzek
8992fbe434
zml.Shape.assertTags (#1022)
we often use `withTags` to document the current shape, but it could also
hide bug by overriding previous tags.
I introduce `assertTags` that should be favored when possible
2026-11-04 16:04:18 +01:00
Guillaume Wenzek
17ba49cce8
zml.ops.manualComputation: fix error message (#1021)
use `.format` instead of default Zig formatting
2026-11-04 16:04:18 +01:00
Guillaume Wenzek
350c6c1606
zml.Mesh: axis/hasAxis helpers (#1019) 2026-11-04 16:04:17 +01:00
Guillaume Wenzek
4707e58774
zml.Compiler: allow to propagate tensor store names (#1018)
The name can appear in the IR and thanks to #1003 it shows in the web
view of the model making all compute blocks pretty obvious.

The screenshot below without names would be quite confusing, 
but now we can see this is the QKV of first layer of a Llama model.
<img width="2211" height="1208" alt="image"
src="https://github.com/user-attachments/assets/1917edb0-a772-4a1d-9aa7-ad5aaf7b5260"
/>
2026-11-04 16:04:17 +01:00
erwanvivien
194175b36d
zml/kernels/flash_attn: add prefill flash_attention kernel and add Causal (#1000)
Prefill kernel is based on libflashattn.so from zml/flash-attention
2026-11-04 14:35:55 +01:00
Hugo Mano
6a1698f270
zml/ops: add manualComputation2 with explicit partitioning (#1013)
Accept a logical mesh and typed input/output partition specs at manual
computation boundaries.

Will replace `zml.ops.manualComputation` in a future PR.
2026-11-03 18:22:59 +01:00
Hugo Mano
4c355c61b7
zml/sharding: keep unconstrained graph results open (#1004)
Ordinary graph results should have an unspecified layout until
constrained or inferred by the partitioner.

Previously, we could report replication simply because a tensor had no
mesh, even when its stored partition specs were open. Other paths
preserved selected operations’ sharding metadata, making the frontend’s
interpretation depend on how a tensor was produced.

This change gives those results a consistent open state while requiring
concrete layouts at executable boundaries.

Shardy explicitly defines the intended semantics: [“no sharding
attribute on a tensor is equivalent to a fully open tensor
sharding.”](https://openxla.org/shardy/sharding_representation#open)

Open dimensions remain available for further partitioning; closed
dimensions fix the specified layout.
Replication is therefore a constraint, whereas an open layout lets the
partitioner choose either sharding or replication:

This distinction also applies to unused mesh axes. Marking an axis
explicitly replicated prevents propagation from using it, even when a
tensor dimension is open. For example, `[{?}, {}]` must leave unused
axes available; adding replicated={"data", "model"} removes those
choices.

Keeping intermediates open lets the partitioner use constraints from
producers and consumers, with operation-specific rules for
transformations such as dot and reshape.
2026-11-03 18:11:15 +01:00
Corentin Kerisit
439ba6c0a6
bazel: register runfiles initialization from C (#1005)
ZML initializes runfiles before `main` so its loaders can locate
packaged libraries and data.

When using `-fno-llvm -fincremental` and linking with the incremental
ELF linker, declaring the registration pointer in Zig produces incorrect
`.init_array` data.

See https://codeberg.org/ziglang/zig/issues/37453

Use a C constructor on Linux and macOS so the C compiler emits the
platform-specific startup registration, while Zig exports the existing
runfiles initialization function directly.
We can revert when https://codeberg.org/ziglang/zig/issues/37453 is
fixed.

With this fix, we bypass the wrong behavior in incremental ELF linker
and get closer to ZML compliance with zig incremental using their
incremental linker.

The registration is shared by the `use_cc_common_link=true` and `false`
paths. Keep the explicit linker retention so the constructor is included
even when its object is in a static archive.
2026-11-03 14:45:23 +01:00
Benjamin Rabier
86ab457d0b
dsl/cute: add SM100 primitives and CTA-group support (#1007)
Add helpers for TMA gather/prefetch, weight-stationary MMA,
shared-memory descriptors, packed FP32 arithmetic, async cluster stores,
and register limits.

Make CTA groups and barrier scopes explicit across TMA, block-scaled
MMA, TMEM, and synchronization, including two-CTA operations and
multicast completion. Consolidate named CTA barriers and update the MoE
and paged-attention callers.

Cover instruction operands, descriptor encoding, and MLIR round-trips in
builder tests.
2026-11-03 10:26:31 +01:00
Louis Lechevalier
2532589572
zml/moe: CuTe MoE nvfp4 on sm103 (#981)
This PR port the CuTe MoE for nvfp4 x nvfp4 from cutlass flashinfer into
zml. It uses the CuTe zig DSL. It is the first PR of a serie that aim to
replace the covering of flashinfer cutlass in ZML. The final goal is to
remove the flashinfer cutlass binding. It begins by removing the binding
of the bf16 kernel on sm90 (never used, we have triton).
2026-11-02 14:46:41 +01:00
Benjamin Rabier
d4b413f5f9
zml/tokenizer: support Mistral's numeric lookahead (#997)
Mistral Large 4's Tekken tokenizer uses positive lookahead to split
number runs into groups of three from the right. IREE's regex compiler
rejects this pattern, preventing the tokenizer from loading.

Patch IREE to recognize the exact isolated Split pattern, match complete
number runs with its DFA, and emit the original groups using Unicode
codepoint counts. Preserve progress when output fills during streaming
and finalization. Register the patch in ZML's IREE repository rule.

Treat RESOURCE_EXHAUSTED as normal encoder backpressure instead of
logging errors before retrying. Add regressions for numeric token IDs,
Unicode numbers, byte-at-a-time input, and output overflow.
2026-10-29 14:10:32 +01:00
Raphael Nahoum
ca39742f14
zml/platform: bump plugins and preallocate Metal memory again (#1001)
- Turns preallocation back on by default on Metal, as before #979: not
preallocating seems to cause memory issues we haven't looked into yet.
- Bumps the PJRT plugins to `202610290951.309.1.785550677e75`.
2026-10-29 12:52:21 +01:00
Raphael Nahoum
e4e031b2c5
zml/kernels/paged_attention/metal: port the kernels to the Metal DSL (#999)
Moves Metal paged attention from the `zml$paged_attn` custom call to
Metal DSL ports of the same kernels.

- `fa_vec_paged`, with a split-K reduce, handles causal bf16 decode.
`paged_attention_tiled` and `zero_padding_rows` handle everything else.
- The shared paged attention test now runs on Metal. Triton counts as
available only on CUDA, ROCm and oneAPI. StableHLO is the reference on
Metal for the causal case.
2026-10-29 12:52:20 +01:00
Raphael Nahoum
70f1bd4af4
dsl/metal: add the Metal DSL (#989)
Adds a Zig DSL for Metal kernels over the AIR dialect, and moves the
Metal paged attention kernels onto it.

- `mlir/dialects/air`: a `thread` address space. Each op has an
`llvmBuilder` that calls its `air.*` builtin with Apple's attributes,
and the dialect's translation interface builds the `!air.kernel`
metadata, so lowering to LLVM IR is MLIR's own translation. The verifier
rejects what Apple's compiler can't take: bool, scalable and n-D
vectors, buffers of non-Metal scalars, scalar f32 converts, and
`air.kernel` on a declaration.
- `dsl/metal`: an MSL-shaped builder. It emits `air` plus the
`arith`/`cf`/`vector`/`llvm` IR that clang produces for MSL, and covers
MSL's math, SIMD-group, barrier and `simdgroup_matrix` functions.
- `zml.kernel.metal` / `ops.metal`: wrap a kernel into a
`__gpu$xla.gpu.air` custom call. XLA lowers it, compiles it to a
metallib and launches it.
2026-10-29 12:52:20 +01:00
Damien Mehala
b725ff2481
zml/Compiler: add XLA flags to improve generated kernels for ROCm (#998)
Enable two flags for ROCm:
- multi output fusion: allows to search for fusion opportunity
generating different output.
- latency hiding scheduler: overlap async communication with
computation.
2026-10-29 11:16:03 +01:00
Damien Mehala
47e692c300
platforms/oneapi: bump to IGC 2.41.15 (#996)
Bump to resolve a compiler bug encountered on LFM3.

LFM3 broadcasts one sigmoid gate per attention head. This generated code
of sigmoid introduced a BF16 conversion and conditional select. The
compiler then lower it to a scalar computation which had a
register-aligment bug. The fix introduced in 2.41.11
<5d92530a07>
enforces alignment.
2026-10-29 11:15:39 +01:00
Raphael Nahoum
b37ba31500
mlir/dialects: build every remaining op and attribute through the dialect packages (#993)
The DSLs and the kernels no longer make ops or set dialect attributes by
name: each dialect package gets the builder or helper they were missing,
and every consumer goes through it. This now fix ZML core files as well
2026-10-27 17:53:31 +01:00
Raphael Nahoum
3736f3dc29
dsl: unify the kernel DSLs (#991)
Makes kernels written in the DSLs read alike, without changing what they
emit.

- Every DSL target is `//dsl/<name>:builder`, and `dsl/common` is
imported as common.
- Scopes without values end with `close()`.
- The scf DSLs (triton, mosaic, cute, fly) share
`openFor`/`openIf`/`openIfElse`/`openWhile` in `dsl/common`. cuda_tile
keeps its own scopes with the same shapes.
- Scopes take tuples or arrays, and result types can be dtypes (.f32),
Values, or MLIR types.
- Typed constants are `constant(dtype, value)`, alongside constIndex and
constantLike.
- Comparisons and bitcasts have the same names in every DSL: Fly gains
`lt`/`eq`/…, and every DSL spells it `value.bitCast(dtype)`.
- One block stack and one `common.verify`, which logs the module that
fails to verify.
- READMEs updated.
2026-10-27 17:53:31 +01:00
Damien Mehala
aa30110932
platforms/rocm: default rocprofiler log level to error (#992)
By default rocprofiler log level is `warning`, which can produce
unnecessary log output. Now, we override this log level to error, unless
`ROCPROFILER_LOG_LEVEL` environment variable is already set.
2026-10-27 13:01:40 +01:00
Kevin Barre
8f40b8b09f
zml/kernels: add TCL paged attention and MoE backends for Furiosa (#973)
zml: add TCL kernels and Furiosa attention and MoE backends

Introduce a typed TCL kernel DSL and MLIR bindings, and select TCL
automatically for Furiosa paged attention and BF16 SwiGLU MoE.

- Support native kernel calls, fusion, in-place outputs, device loops,
  compiler configuration, and chip-aware layouts.
- Implement standard paged attention for decode, prefill, and mixed
  batches, including causal/non-causal masks, sliding windows, custom
  scales, and per-head sinks.
- Use BF16 contractions with FP32 normalization and accumulation.
  Exclude unused KV entries before contractions and return zero for
  inactive or fully masked queries.
- Reuse prepared metadata for the optimized decode path and process
  longer contexts in page-aligned chunks.
- Add activation-stationary and blockwise fused-expert kernels.
- Fix StableHLO attention metadata for variable batch sizes and add
  host-reference, emission, and hardware regression coverage.
- Bump the Furiosa sandbox to 202610221403.283.1.d9f4f68759e1.

---------

Co-authored-by: Raphael Nahoum <raphael@zml.ai>
2026-10-26 15:47:48 +01:00