72263aa9e3
workspace: fix mistakes in modules bumps
...
Fix missing = in stbalehlo integrity.
Changes back - by . in xla module name and folders.
Correctly depend on `xla@20250204.1-6789523`
2024-06-06 09:56:17 +00:00
f7450a2104
stablehlo: bump to head and use new dialect capi
...
This drastically reduce the number of build steps (from 3589 to 2553 steps)
2024-05-31 13:02:46 +00:00
221ece647d
zml/ops.zig: Added zml.ops.case operation
...
This can be used to select which branch will be run at runtime.
It wraps the `stablehlo.case` operation.
2024-05-30 14:11:08 +00:00
27aabf9beb
Add Bazel build rules and a test for the benchmark, llama, mnist, and simple_layer examples.
2024-05-23 15:52:34 +00:00
3aac788544
Update Bazel build configurations (zig.bzl, BUILD files) for MLIR, PJRT, Neuron, ROCm, tokenizer, and tools, fixing broken dependencies.
2024-05-20 11:28:25 +00:00
05944b5cc9
Update FnCache to copy and reuse non‑tensor fields in fixed‑size structs, preventing undefined memory in core modules.
2024-05-15 17:54:52 +00:00
dfe55b0d34
Update Bazel lock file for examples to reflect FnCache non‑tensor handling changes.
2024-05-13 16:59:37 +00:00
8d795dd676
pjrt: profiler support std writer API
...
expose a more low-level function to customize where to write profile
reports
2024-05-09 11:09:29 +00:00
26558d6201
Update examples MODULE.bazel and lockfile to use XLA 20250204.0-6789523 and ensure Bazel 8 compatibility.
2024-05-08 14:03:45 +00:00
f5ab6ff2c6
Update XLA to version 20250204.0-6789523 and adjust Bazel module and runtime files for Bazel 8 compatibility.
2024-05-03 15:57:56 +00:00
a34190679b
Fix llama token handling and remove redundant prompt token reuse in core Zig modules (aio, module, nn, pjrtx, tensor)
2024-05-02 17:10:11 +00:00
394e63e273
Fix llama example to correctly handle token output and avoid re‑feeding the last prompt token.
2024-04-24 16:44:25 +00:00
5a2171793d
workspace: MODULE.bazel cleanup
...
Title says it all !
2024-04-22 09:27:44 +00:00
bafe13f546
Update examples/MODULE.bazel.lock to reflect libxev version bump.
2024-04-18 12:53:16 +00:00
65c28111a9
Update libxev to version 20252401.0‑31eed4e and apply patches and.
2024-04-15 13:03:25 +00:00
13eff4e661
pjrt,zml: add memory bindings
...
This preliminary PR binds PJRT memory endpoints and adds them to
`zml.Buffer`.
A follow up PR will properly integrate it inside `zml.Buffer`
2024-04-11 15:43:24 +00:00
190c6978d2
llama: simplify llama3 prompt template encoding by removing redundant newline re-encoding and ensuring a trailing newline.
2024-04-10 09:36:28 +00:00
d4db5ccc6b
Integrate TinyLlama support, restore the homemade tokenizer, and align Zig API naming across stdx and zml tokenizer modules.
2024-04-05 15:07:29 +00:00
b67685b941
Add example Bazel build files and tokenizer test for tinyllama, including tigerbeetle integration and flags.
2024-04-01 17:40:18 +00:00
567210d1d7
bazel: depend on prebuilt protoc binaries to eliminate ~1300 build steps. Note: integration is currently blocked due to version constraints in rules_proto and toolchains_protoc.
2024-03-29 09:54:57 +00:00
e0c8eecb79
bazel: use OID as sha256 for Git LFS files to prevent unnecessary HuggingFace redownloads.
2024-03-28 17:52:52 +00:00
a811b2e1e3
llama: fix dimensions and data types
...
Removed unnecessary batching dimension introduced by recent changes. Converted index outputs from i32 to u32 for token indices. Ensures Llama runs on CUDA and RoCM. Tested on CUDA.
2024-03-20 13:37:19 +00:00
602757e7a9
Update examples to use the corrected logFn API.
2024-03-18 13:11:14 +00:00
754656f2f0
Replace real mutex with async Mutex for logFn, add fallback logger support outside coroutines, and fix ResetCondition handling.
2024-03-14 11:43:33 +00:00
980f1b17fb
Ensure all runtime plugins have correct SONAME values, fixing issues with prebuilt PJRT plugins.
2024-03-11 10:15:22 +00:00
8a25b1eb74
Revert CUDA PJRT plugin version to 0.4.38 to address performance regression on XLA master.
2024-03-05 17:04:42 +00:00
76e314db9b
Update Llama example docs and Bazel build files, and add tests for the new HuggingFace tokenizer integration.
2024-03-04 12:11:13 +00:00
959bc48c42
Add HuggingFace tokenizer bindings and SentencePiece integration; update BUILD files, async utilities, and FFI modules to support the new tokenizers.
2024-02-28 15:47:37 +00:00
5048e7dc89
Update example lock file for rules_distroless 0.4.2 upgrade and verify MNIST image build works.
2024-02-26 15:30:13 +00:00
b4b2490690
Upgrade rules_distroless to 0.4.2 in MODULE.bazel and refresh MODULE.bazel.lock accordingly.
2024-02-21 17:48:10 +00:00
c109b12e1b
Various minor fixes: rewrite tinyllama tokenizer newline token, prevent HostBuffer.isContiguous false trigger on 1‑dim axes, improve HostBuffer.slice1d error messages, simplify module.zig output to show .mlir file path, correct setFlags handling of comptime int/float, make tokenizer.zig return <oob> for out‑of‑range detokenization, and speed up Buffer.constant creation up to 2.5 GB/s on CUDA.
2024-02-19 12:34:18 +00:00
3970df5b48
Update getting_started tutorial and example Bazel files for Bazel 8 migration.
2024-02-14 10:44:47 +00:00
169a24307c
Migrate workspace and XLA module definitions to Bazel 8, updating MODULE.bazel files, BUILD rules, and related migration patches.
2024-02-12 12:43:23 +00:00
7e6103d876
Upgrade XLA to version 20250122.0-cc075be, switch to nvptx compiler and nvlink with nvjitlink support, add warning for CUDA path in LD_LIBRARY_PATH, and revert the previous CUDA sandbox fix.
2024-02-06 09:31:48 +00:00
b8a0aaee5a
Update tokenizer to handle byte_fallback for Llama3 GPT2 vocab and add a Llama3‑specific normalizer; adjust tinyllama.zig and hostbuffer.zig to use the new tokenization logic.
2024-02-05 15:22:44 +00:00
b643f7bc53
Add Bazel build rule and test for Llama3 tokenizer’s byte fallback and unknown token handling.
2024-02-02 10:25:48 +00:00
5120fe00dc
Update libxev epoll patch to resolve crashes and hangs in epoll and kqueue implementations.
2024-01-29 17:15:11 +00:00
edc2ac26f8
Adjust ROCm runtime sandboxing to hook only the PJRT plugin and make hipblastlt bytecodes optional.
2024-01-26 13:02:23 +00:00
0ce36599da
Update example build config and Llama demo to support the new async epoll backend and zigcoro scheduler.
2024-01-22 12:17:01 +00:00
a7b7ae0180
Fix async hangs by reworking the libxev epoll backend and using callBlocking for PJRT plugin loading, improving performance across async and runtime modules.
2024-01-16 14:13:45 +00:00
434cee3a6c
Fix CUDA and ROCm sandbox discovery, update epoll libxev patch to prevent high CPU usage, enable XLA GPU latency‑hiding scheduler, and upgrade cuDNN to 9.6.0.
2024-01-15 09:41:42 +00:00
5b8e42f9a9
Vendor zigcoro and unify APIs; rework internals for stdx.meta compatibility, add Channel.try_send/try_recv methods, support dynamically sized channels with comptime capacity, and introduce PoolStackAllocator for coroutine stack allocation.
2024-01-11 15:40:15 +00:00
68dbc290e9
zml: revamp scatterSlices
...
Main issue with current `scatter` implementation is that it uses
broadcasting dims of `stablehlo.scatter`.
While nice in theory, the optimizer doesn't handle them well and they
often are unrolled into while loop.
Here I convert the batching dim to extra iotas indices.
2024-01-08 17:55:20 +00:00
83b5e1ec48
fix
...
Before we where using `module.op().writeBytecode(writer)` to compute the
hash of a model
but it crashes on some inputs, notably for unused variables.
So I used the text representation of the mlir.
2024-01-05 16:44:41 +00:00
acc492454f
Add operator name to source locations and introduce QoL enhancements: remove bias from sdpa, support shape literals in gatherSlices, add Shape.outer, Tensor.all, and infer argMax dtype.
2024-01-01 15:31:41 +00:00
223857251d
Update MNIST example to use new operator source locations and reflect recent API changes (sdpa bias removal, gatherSlices shape literals, Shape.outer, Tensor.all, and argMax dtype inference)
2023-12-26 10:45:52 +00:00
5bd7f8aae9
zml: HostBuffer.prettyPrint()
...
Add pretty printing of HostBuffer.
This will be leverage by the debug helper `x.print()`
It can also be used like this: `std.log.info("my buffer: {}",
.{host_buffer.pretty()})`
2023-12-25 13:01:17 +00:00
5ddd034d2c
pjrt: Fix profiler by allowing i64 resource IDs and reserving memory when creating array lists.
2023-12-20 17:18:02 +00:00
7ef87236ce
Rewrite simple transpose as reshape in core ZML modules and raise default profiler event limit to 1,000,000.
2023-12-18 13:56:45 +00:00
8a031bd4c8
Update Llama example to use the simplified transpose implementation and increase default profiler size to 1,000,000 events.
2023-12-15 12:06:42 +00:00
