Fastllm Triton Ops

FastLLM Triton Ops

Overview

Use the existing Linear Triton prototype as the reference architecture: C++ decides whether an op is eligible, starts a local Python compiler server only on demand, asks it to emit a cached cubin plus metadata, then launches that cubin through the CUDA Driver API. Every Triton path must be environment-gated and must fall back to the original CUDA implementation on unsupported inputs or compile/launch failure.

Workflow

1. Inspect the existing CUDA op path in src/devices/cuda/cudadevice.cpp.

   • Find the op's Reshape, CanRun, Run, and lower-level helper functions.
   • Identify the exact tensor layout, dtype combinations, shape variables, optional bias/scale tensors, and output aliasing behavior.
   • Keep the original implementation as the fallback path.

2. Add or extend the Python compile path in tools/fastllm_triton_server.py.

   • Add a @triton.jit kernel for the new op.
   • Add <op>_cache_paths(payload) with a deterministic filename that includes op name, dtype/layout variants, SM arch, compile-time tile sizes, and feature flags.
   • Add compile_<op>(payload) that validates payload fields, compiles with ASTSource, writes .cubin, writes .json metadata, and returns the metadata.
   • Extend handle_compile(payload) by dispatching on payload["op"].
   • Keep /health and /compile stable; do not break existing "op": "linear" requests.

3. Add a CUDA launch wrapper in src/devices/cuda/fastllm-triton-cuda.cu.

   • Reuse LoadTritonKernel for cubin/module/function caching.
   • Add one extern "C" wrapper per op, for example FastllmCudaTriton<Op>(...).
   • Use FastllmCudaPrepareInput, FastllmCudaPrepareOutput, FastllmCudaFinishInput, and FastllmCudaFinishOutput when passing Data buffers.
   • Match the Triton kernel argument order exactly, including Triton's hidden global_scratch and profile_scratch pointer arguments when needed by AOT metadata.
   • Return false on load or launch failure so the C++ caller can fall back.

4. Declare the wrapper in include/devices/cuda/fastllm-cuda.cuh.

   • Keep the signature close to the CUDA fallback helper's shape arguments.
   • Pass metadata fields needed for launch, such as kernelName, shared, numWarps, and tile sizes.

5. Wire the op in src/devices/cuda/cudadevice.cpp.

   • Add a small metadata struct for the op's .json fields.
   • Reuse common helpers: CudaTritonCacheDir, CudaTritonDataTypeName, CudaTritonHttpRequest, and CudaTritonEnsureServer.
   • Add CudaTriton<Op>BaseName, CudaTritonRead<Op>Meta, CudaTritonRequest<Op>Kernel, and TryCudaTriton<Op>.
   • Gate with global FASTLLM_CUDA_TRITON; add an op-specific override such as FASTLLM_CUDA_TRITON_<OP>=0.
   • Validate device, pointer presence, dtype, layout, shape, arch, and feature constraints before requesting a kernel.
   • In the original op helper, call TryCudaTriton<Op>(...) immediately before the original CUDA implementation.

6. Update build wiring only when needed.

   • src/devices/cuda/fastllm-triton-cuda.cu is already in CMakeLists.txt.
   • If adding new files, update CMakeLists.txt and link dependencies without changing unrelated targets.

Current Contract

The existing Triton infrastructure uses these environment variables:

• FASTLLM_CUDA_TRITON=1: enable Triton-backed CUDA ops globally.
• FASTLLM_CUDA_TRITON_<OP>=0: disable one op while keeping the global flag on, for example FASTLLM_CUDA_TRITON_LINEAR=0.
• FASTLLM_CUDA_TRITON_CACHE_DIR: override cubin/json cache directory.
• FASTLLM_CUDA_TRITON_SERVER_HOST, FASTLLM_CUDA_TRITON_SERVER_PORT: choose compiler server endpoint.
• FASTLLM_CUDA_TRITON_PYTHON: choose Python interpreter.
• FASTLLM_CUDA_TRITON_SERVER_SCRIPT: choose server script path.
• FASTLLM_CUDA_TRITON_SERVER_LOG: choose compiler server log path.
• FASTLLM_CUDA_TRITON_SERVER_WAIT_MS: choose startup wait timeout.
• Per-op tile knobs should use FASTLLM_CUDA_TRITON_<OP>_<PARAM>, matching Linear's BLOCK_M, BLOCK_N, BLOCK_K, NUM_WARPS, and NUM_STAGES.

The metadata JSON returned by the compiler server should include at least:

{
  "ok": true,
  "op": "op_name",
  "cubin": "/path/to/kernel.cubin",
  "kernel": "compiled_kernel_name",
  "shared": 0,
  "num_warps": 4
}

Add op-specific launch fields, such as tile sizes, only when the C++ launcher needs them.

C++ Pattern

Keep the C++ control flow shaped like this:

if (!CudaEnvFlagEnabled("FASTLLM_CUDA_TRITON")) {
    return false;
}
const char *opEnv = std::getenv("FASTLLM_CUDA_TRITON_MYOP");
if (opEnv != nullptr && opEnv[0] != '\0' && !CudaEnvFlagEnabled("FASTLLM_CUDA_TRITON_MYOP")) {
    return false;
}
if (!inputs_are_supported) {
    return false;
}

Meta meta;
if (!ReadMeta(metaPath, meta)) {
    if (!RequestKernel(..., meta)) {
        return false;
    }
}
return FastllmCudaTritonMyOp(meta.cubinPath.c_str(), meta.kernelName.c_str(), ...);

Do not throw or call ErrorInFastLLM from the Triton trial path unless the original CUDA path would also fail. Unsupported Triton cases should return false.

Validation

Run validation in layers:

1. Build:

bash install.sh -DUSE_CUDA=ON

2. Unit or op test, with Triton enabled and an isolated cache:

FASTLLM_CUDA_TRITON=1 \
FASTLLM_CUDA_TRITON_CACHE_DIR=/tmp/fastllm-triton-optest \
../optest --op linear --device cuda:0 --param batch=4 --param in=8 --param out=6

Adjust the optest command for the op being added.

3. End-to-end server smoke test:

FASTLLM_CUDA_TRITON=1 \
FASTLLM_CUDA_TRITON_CACHE_DIR=/tmp/fastllm-triton-qwen \
ftllm server ~/hfmodels/Qwen3-8B/ --device cuda:0 --host 127.0.0.1 --port 18080 --tokens 8192 --hide_input

Send one non-streaming request to /v1/chat/completions and confirm it completes.

4. Benchmark after warmup.
   • Always exclude first compile/server startup from performance numbers.
   • Compare against the same command without FASTLLM_CUDA_TRITON.
   • Report token throughput and wall time; state whether the measurement is kernel-only or end-to-end server throughput.

Guardrails

• Keep Triton optional: default behavior must stay unchanged when FASTLLM_CUDA_TRITON is unset.
• Prefer compile keys that are independent of runtime shapes when the kernel supports dynamic dimensions.
• Include every compile-time specialization in the cache filename to prevent stale cubin reuse.
• Serialize compilation in the Python server with the existing lock unless proving concurrent compilation is safe.
• Keep C++ JSON parsing defensive; missing or invalid metadata should fall back.
• Clean up test servers and compiler-server processes after benchmarks.