wasmgpu

Run independent WebAssembly instances on a hardware GPU from Python 3.8+.

import wasmgpu

module = wasmgpu.Module("worker.wasm")
with module.spawn(100_000) as instances:
    results = instances.call("process", inputs)

The runtime is implemented in this repository:

Python → WASM validation / specialized WGSL → wgpu-py → Metal / Vulkan / DX12

Python parses and validates the module, uploads initial state, dispatches work, and reads results. All guest instructions and WASI services run on the GPU. There is no CPU interpreter, CPU fallback, host WASI service loop, or dependency on Wasmtime in the installed package. Software GPU adapters are rejected.

execution="auto" (the default) and execution="compiled" compile every supported function. The GPU program buffer contains metadata, jump tables and embedded data, no instruction bytecode. Execution shaders do not contain an opcode decoder or interpreter. execution="interpreter" explicitly selects the original GPU interpreter for differential testing; it is never an implicit fallback. A compilation failure is reported to the caller.

The compiler divides the complete control-flow graph into bounded units, including functions larger than one unit. Native pipelines are compiled lazily when their continuations are reached. Limits of 256 lowered instructions and 64 KiB of generated source apply per unit, not per module; source overflow splits a unit instead of dropping functions. Large br_table targets are ordinary jump-table data. The driver compiles generated WGSL into native GPU code through wgpu. Python 3.8 / wgpu-py 0.18 uses a 16 KiB source cap per unit to reduce pressure on its older Metal compiler.

Basic blocks keep intermediate values and modified locals in shader variables, spilling at continuations or traps. Calls and recursion use explicit GPU frames; there is no WGSL recursion. A dispatch loop selects compiled block addresses, never opcodes. When fuel or stack space cannot accommodate a whole block, an unrolled, statically compiled prefix preserves the exact trap and prior effects. It does not use the reference interpreter. Numeric implementations include software f64 and checked word loads. WASI services use a separate shared GPU pipeline, preserving the same embedded filesystem and operand stack. Python schedules pipelines without executing guest instructions or WASI operations.

This architecture removes instruction interpretation, but transitions between compilation units require GPU scheduling and synchronization. Lazy native compilation can dominate a first invocation; measurements must separate it from execution. It does not establish a speedup over Wasmtime on CPython workloads. Use the independent development watchdog below for GPU experiments. Source-size limits alone do not guarantee driver memory usage or compilation time.

module = wasmgpu.Module("worker.wasm", execution="compiled")
with module.spawn(1000) as instances:
    results = instances.call("process", inputs)
    print(instances.last_call)  # phase timings and compiled/interpreted counts
    instances.reset()          # fresh guest state, same allocated GPU buffers

compile_functions=[...] optionally prioritizes function indices when grouping units; every other function remains compiled. compile_limit (1–256) may lower the per-unit instruction budget. Parsed modules, generated code and GPU pipelines have bounded in-process caches. Instances share an immutable initial template; batched heap copies and per-instance random nonces are initialized on the GPU. reset() restores memory, globals, tables, files and descriptors and reruns the core WASM start function. It does not call a WASI command's _start.

This is an experimental engine with the supported profile below, not a complete implementation of every WebAssembly proposal or a security sandbox. Apple M4 / Metal has been tested. Vulkan and DX12 use the same shader but have not yet been verified on physical hardware in this project.

Use a virtual environment for installation, development and tests:

python3 -m venv venv
venv/bin/python -m pip install -e .
venv/bin/python -m pip install -r requirements_dev.txt

On Windows, use venv\Scripts\python.exe instead. Dependency markers select wgpu-py 0.18.0 for Python 3.8, 0.24.0 for 3.9, 0.31.0 for 3.10, and 0.32.0 for 3.11+. Actual Metal execution was verified with Python 3.8, 3.9, 3.10 and 3.11. The other Python/backend combinations still require hardware testing.

Module accepts a path or binary WASM bytes. It does not compile Rust, C or WAT. module.exports lists export names and kinds. Instantiate once, then reuse the instances: linear memory, globals, tables, files and descriptors persist between calls. Instances have independent state, including across internal GPU batches.

with module.spawn(3) as instances:
    a = instances.call("one_argument", [10, 20, 30])
    b = instances.call("two_arguments", [(1, 2), (3, 4), (5, 6)])
    c = instances.call("no_arguments")

There must be exactly one input row per instance. Input validation completes before any instance executes. Results are a list in instance order: scalars for one result, tuples for multiple results, and None for no results. Integers return as signed i32/i64; integer inputs accept signed or unsigned bit patterns. Floating inputs/outputs are Python floats; Python may quiet signalling f32 NaNs at this API boundary. Only null references can be supplied from Python.

instances.write_memory(1024, b"payload", instance=0)
content = instances.read_memory(1024, 7, instance=0)

These are explicit data transfers. Reading or writing memory does not execute guest code on the CPU. A core WASM start function executes during spawn. For WASI reactor modules, call an exported _initialize once if the compiler requires it; for command modules, explicitly call _start.

Trap provides traps (instance index → reason) and results (including successful peers). Completed effects before a trap persist; calls are not transactions. A later call can reuse the instances. proc_exit raises a trap and records the per-instance code in instances.exit_codes. Explicitly close instances, preferably with a context manager, to release GPU buffers.

call(..., cancel=callback) checks cancellation between GPU dispatches and raises InterruptedError if the callback returns true. Already completed effects persist. It cannot preempt a GPU command already submitted to the driver. last_call records lazy WGSL generation and native compilation, input preparation, upload, dispatch/completion synchronization, result readback and decoding separately, including counters up to cancellation. In compiled mode, last_call.interpreted_instructions is always zero.

Files are byte contents embedded into each instance's GPU filesystem. There are no host directory mounts, host file operations or network access.

module = wasmgpu.Module("worker.wasm", files={"data/input.txt": b"1.25\n2.5\n"})
wasi = wasmgpu.Wasi(
    args=["worker", "data/input.txt"],
    env={"MODE": "batch"},
    stdin=b"input stream\n",
    storage_size=256 * 1024,
    max_files=64,
    max_fds=64,
    seed=123,
)
with module.spawn(8, wasi=wasi, memory_pages=32, stack_size=4096) as instances:
    instances.call("process")
    output_file = instances.read_file("result.txt", instance=0)
    stdout = instances.stdout  # list of captured bytes, one per instance
    stderr = instances.stderr

Wasi(files=...) overrides same-named Module(files=...) entries. Configuration and contents are copied at instantiation. Root is preopened at fd 3 as .; fd 0/1/2 are emulated stdin/stdout/stderr. Paths use /, are limited to 255 UTF-8 bytes after resolution, and cannot escape their directory capability. max_files includes directories, symlinks and four reserved entries; storage_size includes stdin, stdout, stderr and all file contents.

The shader implements file creation, reads/writes and positioned I/O, seeks, truncation/allocation, descriptor rights, metadata/timestamps, directory enumeration, rename, hardlinks, symlinks and unlink. Deleted storage is reclaimed; open descriptors and hardlinks keep their inode alive. read_file is an inspection API for a regular file's stored path; guest code resolves symlinks.

Other operating-system services have explicit virtual semantics:

• Arguments/environment come from the embedded configuration.
• All clocks use a per-instance virtual counter. It starts at clock_epoch_ns (default 0) and advances by clock_resolution_ns (default 1) per lowered WASM instruction. It never reads the host clock.
• poll_oneoff reports ready virtual descriptors or advances virtual time to the earliest clock deadline, without sleeping on the host.
• random_get uses ChaCha20 on the GPU. The seed is a nonzero u32 or a 32-byte key; an instance index supplies its nonce. Streams are deterministic and independent of batch size. The default seed is public and provides no unpredictable system entropy.
• sched_yield is a no-op in the isolated instance model. Signals terminate the affected invocation; no host process is signalled.
• No virtual sockets are provisioned. Socket imports return BADF for invalid descriptors and NOTSOCK for existing non-socket descriptors. They never open host sockets. sync/datasync operate on the in-memory filesystem only.

All 46 Preview 1 imports have signature validation and GPU dispatch. This is an emulated environment, not a promise of an ordinary operating system or complete WASI conformance. Preview 2/3 and arbitrary host imports are unsupported.

Supported: all scalar MVP numeric instructions; i32/i64; software IEEE-754 f32 and f64 including subnormals, signed zero and rounding; direct/indirect recursive calls; blocks/loops/branches; multi-value; mutable globals; memory32; multiple funcref/externref tables; reference instructions; sign extension; saturating conversions; bulk memory/table operations and passive/declarative segments.

Currently unsupported: SIMD, threads/shared memory, exceptions, tail calls, GC, typed function references, memory64, multiple linear memories, imported memories/tables/globals, and module linking. Unsupported features fail explicitly. They never trigger execution through a CPU engine. Tables and linear memory have fixed GPU growth budgets; grow returns -1 when the reserved capacity is reached. The current memory32 addressing implementation caps memory at 65,535 pages.

spawn exposes resource controls:

All persistent instance state stays on the GPU. Internal batching respects device buffer limits; it does not page state to a CPU runtime. Thus 100,000 tiny workers are practical, but 100,000 workers with 1 MiB private memory require about 100 GiB before stacks/files. Excessive allocations fail before allocation. resident_bytes and adapter_info expose the allocation estimate and selected hardware.

Compiled dispatches yield at basic-block boundaries; fuel remains exact. A dispatch quantum is not a real-time deadline. Large individual bulk operations and WASI operations can take longer than a scalar instruction. Tune resource budgets for trusted workloads; this runtime is not suitable for hostile modules.

venv/bin/python -m tests.gpu_guard --seconds 600 -- -m pytest tests -q
venv/bin/python -m pytest tests -m 'not gpu' # parser/configuration/failure tests only
venv/bin/python -m tests.gpu_guard --seconds 600 -- -m pytest tests -q --wasmgpu-execution interpreter
WASMGPU_COVERAGE_BRANCH=true venv/bin/python -m tests.gpu_guard --seconds 600 -- -m coverage run -m pytest tests -q
venv/bin/python -m coverage combine
venv/bin/python -m coverage report -m
venv/bin/python -m tests.benchmark           # run alone, without concurrent GPU jobs
venv/bin/python -m build

The GPU suites run on Apple M4 / Metal with Python 3.11. The interpreter was also tested with Python 3.8/3.9/3.10 and their pinned GPU backends; native compilation is additionally checked on Python 3.8 / wgpu-py 0.18. The full suite on older backends still requires verification. --wasmgpu-execution interpreter reruns the existing suites against the reference GPU interpreter. Additional differential tests compare compiled and interpreted continuations, precise fuel exhaustion, WASI clocks, cancellation and reset. Large functions, cross-pipeline recursion, exported imports and large jump tables are covered explicitly. Tests forbid use of the interpreter and inspect shader sources and the uploaded program format. Python coverage does not measure WGSL. Tests compare scalar operations with Wasmtime, exercise 100,003 concurrent instances, and run actual compiled C and Rust fixtures with allocation, internal calls, f64, libc/Rust formatting and embedded files. Forty unmodified official WebAssembly 2.0 core suites contribute over 20,000 module/action/assertion commands, including malformed modules and precise NaN bit checks. See fixture provenance. They are a selected subset, not the complete spec suite; Python coverage does not measure WGSL instruction coverage.

Benchmark script records medians after warmup, hardware and versions in the local tests/benchmark-results.json file. This generated report is ignored by Git and excluded from distributions; use --output to choose a different path. GPU timing includes Python packing, transfers, every dispatch/synchronization and result decoding. Two CPU baselines use Wasmtime: one Python call per worker, and a single WASM loop that writes every output followed by reading those outputs into a Python list. Module compilation/instantiation is excluded from call timings; GPU spawn is separate.

Compare both CPU baselines when interpreting results: reducing Python call overhead alone does not establish an acceleration of WASM computation over CPU JIT execution. Performance depends on the workload, instance count and hardware.

CI runs only Python tests that require no GPU, static checks, and package builds, as configured for ordinary hosted runners. The Linux / Python 3.11 job uploads its coverage report to Coveralls with the python-only flag and saves it as a GitHub artifact. This report does not claim GPU tests have run. Full conformance and performance tests must be run locally with a hardware GPU.

For real CPython WASI and the startup/pyflakes/mypy guest workloads used by throng, use benchmark_cpython.py. Supply a runtime directory containing python.wasm and lib/; linter scenarios also take --wheels with pure-Python wheels (pyflakes 3.3.2, mypy 1.14.1 and dependencies). The benchmark embeds these bytes for GPU execution. Only the Wasmtime test oracle mounts host files. The guest scans project/ instead of . to exclude its embedded runtime; throng scheduling, snapshotting and installation overhead are excluded.

venv/bin/python -m tests.benchmark_cpython --runtime /path/to/runtime --engine wasmtime --scenario startup --output /tmp/cpu.json
venv/bin/python -m tests.gpu_guard --seconds 180 -- -m tests.benchmark_cpython --runtime /path/to/runtime --engine compiled --scenario startup --reference /tmp/cpu.json --output /tmp/gpu.json

Use --engine interpreter for the GPU reference, --scenario pyflakes --wheels /path/to/wheels --files 1 for a linter workload, and --files 100 for its larger corpus. --reference checks the runtime hash, arguments, exit code and captured streams. Failed or interrupted runs return a nonzero status and must not be counted as completed performance measurements. Reports are opt-in local files, not repository artifacts. Each invocation is one observation, not a statistical speedup claim; repeat under the same conditions before drawing conclusions.

--count-fuel instruments Wasmtime to estimate a workload's size; its fuel units are not identical to this engine's lowered-instruction accounting, and the instrumentation changes CPU timings. Large linter workloads require explicit --fuel and --call-seconds settings as well as a matching watchdog deadline. Fuel and execution counters are u64 on the GPU; long invocations do not wrap at four billion instructions. Initial compilation, cached module loading and reset timings are reported separately. Lazy compilation during a call is reported in metrics.codegen_seconds and metrics.compile_seconds; call_seconds includes these phases, while metrics.execute_seconds excludes them.

The watchdog runs separately from the worker. On macOS it accounts for physical memory of the worker and its own Metal compiler services, checks compilation and experiment deadlines, and terminates only those processes on a limit violation. Defaults are 1.5 GiB and 30 seconds per pipeline compilation. The benchmark also has a guest fuel budget and a call deadline. A driver cache hit can make a new process's pipeline creation faster; reported compilation times do not imply a cold Metal driver cache. Dispatch timings include completion-state readback; the output readback phase is measured separately.