# CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

## Project Overview

FlashInfer is a GPU kernel library for LLM serving that uses **JIT (Just-In-Time) compilation by default**. This means kernel code changes are automatically picked up without reinstalling the package - extremely convenient for development.

## Quick Reference

| Task | Command |
|------|---------|
| Install for development | `pip install --no-build-isolation -e . -v` |
| Initialize submodules | `git submodule update --init --recursive` |
| Install CUPTI for benchmarking | `pip install -U cupti-python` |
| Run regular tests | `pytest tests/` |
| Run full parameter matrices | `pytest tests/ --full` |
| Run specific test | `pytest tests/path/test_file.py::test_function` |
| Run multi-GPU test | `mpirun -np 4 pytest tests/comm/test_allreduce_unified_api.py` |
| Run benchmark | `python benchmarks/flashinfer_benchmark.py --routine <name> <flags>` |
| Run linting | `pre-commit run -a` |
| Dump environment report (bug reports) | `python -m flashinfer.collect_env` (or `flashinfer collect-env [--json]`) |
| Install pre-commit hooks | `pre-commit install` |
| Clear JIT cache | `rm -rf ~/.cache/flashinfer/` |
| Inspect devcontainer compiler cache | `sccache --show-stats` |
| Enable API logging (basic) | `export FLASHINFER_LOGLEVEL=1` |
| Enable API logging (detailed) | `export FLASHINFER_LOGLEVEL=3` |
| Enable API logging (with stats) | `export FLASHINFER_LOGLEVEL=5` |
| Set API log destination | `export FLASHINFER_LOGDEST=mylog.txt` |
| Enable verbose JIT logging | `export FLASHINFER_JIT_VERBOSE=1` |
| Enable debug build | `export FLASHINFER_JIT_DEBUG=1` |
| Set target architectures | `export FLASHINFER_CUDA_ARCH_LIST="8.0 9.0a"` |
| Set parallel compilation | `export FLASHINFER_NVCC_THREADS=4` |
| Limit parallel ninja jobs | `export MAX_JOBS=4` |
| Enable GDN native short-T path | `export FLASHINFER_GDN_WY_NATIVE_T=1` |
| Enable GDN strided QKV path | `export FLASHINFER_GDN_WY_STRIDED_QKV=1` |
| Enable GDN native A/B tensors | `export FLASHINFER_GDN_WY_NATIVE_AB=1` |
| GDN ReplaySSM u-cache flush kernel with an fp16 state pool (on SM100 served by the UMMA backend; `gated_delta_rule_mtp_ucache_flush(backend=...)`) | `export GDN_UCACHE_STATE_DTYPE=fp16` |
| Opt in to multi-wave KV split in trtllm-gen GQA decode (long or mixed-length KV) | `export FLASHINFER_TRTLLM_GEN_MULTI_WAVE_KV_SPLIT=1` |
| Let `backend="auto"` pick experimental backends | `export FLASHINFER_ALLOW_EXPERIMENTAL_AUTO_BACKENDS=1` |
| Override ragged-prefill `auto` backend order (Blackwell) | `export FLASHINFER_RAGGED_AUTO_BACKEND_ORDER=cutlass,cudnn` |
| Override CuTe-DSL prefill scheduling | `export FLASHINFER_CUTE_PREFILL_PERSISTENT=0` (non-persistent) or `1` (persistent) |
| Skip MoE EP CuTe-DSL import/version guard | `export FLASHINFER_MOE_EP_SKIP_DSL_CHECK=1` |
| Override MoE EP knob-cache path | `export FLASHINFER_MOE_EP_KNOB_CACHE=/path/to/knobs.json` |
| Disable MoE EP fused staging kernel | `export FLASHINFER_MEGA_FUSED_STAGE=0` |
| Log MegaMoE compilation and workspace diagnostics | `export FLASHINFER_MEGAMOE_LOG_COMPILE_SPEC=1` |
| Enable MoE EP EXPERT_MAJOR padding-row exclusion (opt-in; shape-dependent) | `export FLASHINFER_MOE_EP_EXCLUDE_PADDING_ROWS=1` |
| Enable distribution-aware MoE autotune and kernel dispatch (experimental; TRT-LLM MoE only) | `export FLASHINFER_DIST_AWARE_AUTOTUNE=1` |
| Prims-TS debug checks / task-manager verify | `export FLASHINFER_PRIMS_TS_DEBUG_CHECKS=1` |
| Override Prims-TS CuTe-DSL compile options | `export FLASHINFER_PRIMS_TS_COMPILE_OPTIONS="--opt-level 2"` |
| Configure NVFP4 4over6 quantization (preferred over the env vars) | `nvfp4_quantize(..., nvfp4_4over6=NVFP44Over6Config(e4m3_max=256))` |

The minimum Python version used by CI and build tooling is defined in
`.python-version`. CI Docker images use a stable Conda environment name and
derive their interpreter and site-packages paths from that file. When raising
the minimum, also update `project.requires-python` and `tool.mypy.python_version`
in `pyproject.toml`.

## Quick Start for Development

### Installation

```bash
git clone https://github.com/flashinfer-ai/flashinfer.git --recursive
cd flashinfer
pip install --no-build-isolation -e . -v
```

**Important**: The `--recursive` flag is required to initialize submodules in `3rdparty/` (cutlass, spdlog).

If you forgot `--recursive` when cloning:
```bash
git submodule update --init --recursive
```

That's it! You can now:

- Run all benchmarks and unit tests
- Modify kernel source code in `include/` without reinstalling
- Changes are JIT-compiled on next use

The `--no-build-isolation` flag prevents pip from pulling incompatible PyTorch/CUDA versions from PyPI.

### How JIT Compilation Works

When you call a FlashInfer API:

1. **First call**: Generates specialized CUDA code based on parameters (dtype, head_dim, etc.), compiles it with ninja, caches the .so file
2. **Subsequent calls**: Uses cached compiled module
3. **After kernel changes**: Automatically detects changes and recompiles

**No manual rebuild step needed** - just edit `.cuh` files and run your code again.

### Pre-compiled Packages (Optional)

FlashInfer provides optional pre-compiled packages for users who want faster initialization:

- `flashinfer-jit-cache`: Pre-built kernel cache
- `flashinfer-cubin`: Pre-compiled kernel binaries

**For development, you typically DON'T need these.** JIT compilation is fast enough and gives you live code reload.

### Compiler Caching in Devcontainers

The devcontainers run every nvcc and host C++ compile through
[sccache](https://github.com/mozilla/sccache) by setting `FLASHINFER_NVCC_LAUNCHER`
and `FLASHINFER_CXX_LAUNCHER`. Cached objects live in a per-devcontainer Docker
volume mounted at `~/.cache/sccache`, so they outlive a container rebuild — which
wipes `~/.cache/flashinfer` and would otherwise force a full recompile.

```bash
sccache --show-stats                  # hit rate, cache size, uncacheable reasons
export FLASHINFER_NVCC_LAUNCHER=""    # disable for the current shell
```

Caveats:

- Toggling a launcher changes the compile commands in `build.ninja`, so the first
  build after switching it on or off recompiles objects that already exist.
- Link steps, cubin embedding, and CuTe-DSL kernels never invoke nvcc, so they are
  not cached.
- The cache is per-devcontainer by design: sccache's local storage supports a
  single server at a time, so two containers must not share one cache directory.

GitHub PR Test GPU jobs keep a different, per-runner cache: they set
`CI_COMPILE_CACHE_DIR` to a directory under the runner's work directory, and
`ci/bash.sh` namespaces it by runner and the pulled image's immutable ID.
`CI_COMPILE_CACHE_MODE=read-only` (the default) mounts a read-only seed and runs
against a private writable copy, discarded even after test failures. The PR Test
workflow selects `read-write` only for pushes to `main`. A stable per-runner lock
protects maintenance and execution, including accidentally shared workspaces.
Other `ci/bash.sh` callers leave caching unset and run cold.
Before each trusted writable job its runner's cache is cleared when it exceeds
`CI_COMPILE_CACHE_MAX_MB`
(default 40960) or its volume has less than `CI_COMPILE_CACHE_MIN_FREE_MB`
(default 20480) free. Only content-keyed artifacts survive between jobs: CuTe-DSL (with
`FLASHINFER_CUTE_DSL_STRICT_CACHE_KEY=1`), TIRx, KDA PTX, Triton and downloaded
cubins. nvcc modules, tuning results and the generated-source tree are pruned,
since a fresh checkout rebuilds nvcc modules anyway. Each test job's
`FLASHINFER_JIT_BUILD_LOG` is summarized per toolchain by
`scripts/summarize_jit_builds.py` in the job summary.
See [`ci/COMPILE_CACHE.md`](ci/COMPILE_CACHE.md) for lifecycle tests and trust
boundaries.

## cuDNN linear attention

Use `nvidia-cudnn-frontend>=1.31.0`; KDA and GDN normalization use additive
`1e-6` without an epsilon argument. Auto selection also requires CuTe DSL 4.7+.
The bounded KDA SM100/SM103 and GDN SM100 policies live in `kda.py` and
`gdn_prefill.py`. KDA's head/length regions are documented in `docs/api/kda.rst`.
Auto-selected cuDNN ignores the native KDA workspace; native fallback uses it.
Validate output and final state before changing a policy. Explicit state-pool calls
require unique, in-range slots; auto keeps pools on their native route.

Use `benchmarks/bench_cudnn_linear_attention.py` for public-call comparisons.
Its raw-gate GDN baseline includes Torch gate transforms; it is not a vLLM
fused-producer baseline. Keep host CPU, completed wall and reset-inclusive
CUDA graph times separate.

## Testing

Run the regular test suite:

```bash
pytest tests/
```

Run full parameter matrices (used by nightly testing):

```bash
pytest tests/ --full
```

Run specific test file:

```bash
pytest tests/attention/test_hopper.py
```

Run specific test function:

```bash
pytest tests/attention/test_hopper.py::test_single_prefill
```

The sharded test runner defaults to one worker per visible GPU. H100 PR CI opts
into `--workers-per-gpu 2`: eight workers on four GPUs. Other CI lanes retain
the default. Locally, `scripts/task_run_unit_tests.sh --workers-per-gpu 2` enables
sharing; `--workers N` sets the total instead. Workers share the existing JIT
compilation budget.

Scheduling markers apply to the whole source file:

- `node_exclusive`: run alone across the allocation after parallel work, with
  all GPUs visible. Use for multi-GPU requirements or host-memory OOM risk.
- `gpu_exclusive`: run alone on the assigned GPU; other GPUs can keep working.
- `host_memory_heavy`: prioritize this file and serialize its units with other
  host-heavy work, while lighter files can overlap.

Host and GPU reservations cover setup, execution, and process cleanup, and are
acquired together. Blocked units stay queued so workers can take runnable work
from their own queue or steal from another. Pending GPU-exclusive work can
drain one GPU without reserving the host-heavy slot or blocking other GPUs.
`RESOURCE WAIT` reports worker idle time of at least one second while work is
queued but cannot run on that worker's resources.

These markers enforce overlap rules, not memory limits. Most H100 annotations
use measured peaks above 128 GiB host RSS or 32 GiB GPU memory. Attention sink
also uses both narrower markers: its single-GPU H100 run peaked at 89.2 GiB host
RSS and 4.5 GiB GPU memory, and other GPUs can keep working while it runs.
Cold JIT compilation and other GPU families can change the peaks. Validate
completion rate, OOM failures, and repeated wall times with matching scope and
cache conditions when changing a CI lane's concurrency policy.

Capacity reports include the serial floor from host-heavy and node-exclusive
work. They remain optimistic about GPU contention because the estimator uses
worker counts without GPU topology or measured memory capacity.

Run the CPU-only coordinator suite with
`pytest --confcutdir=tests/test_sharding tests/test_sharding/`.

Communication-test wrappers optionally emit cache-publication completion reports
when `UNIT_TEST_RUN_STATUS_DIR` is set to a fresh directory unique to that wrapper
and container/rank. `test_run_status.sh` records each pytest invocation, and
`test_run_status.py` atomically writes `script-result.json` (schema version 1,
`complete`, `exit_code`, `tests_executed`, and `infrastructure_errors`). A small
pytest plugin records execution and collection/internal errors inside each MPI
or torchrun worker; launcher exit 1 alone is not evidence of an assertion failure.
All local worker ranks must finish, and at least one test must execute, to allow
publication. A completed run with assertion failures remains eligible, while
crashes, timeouts, and unfinished invocations do not. With the variable unset,
reporting is disabled. The CPU suite above covers this contract without GPU
dependencies.

### Skipping Tests Based on CUDA Architecture

Use `flashinfer.utils` functions to skip tests on unsupported GPU architectures:

**Available check functions:**
- `get_compute_capability(device)` - Returns `(major, minor)` tuple
- `is_sm90a_supported()` - Hopper (requires CUDA 12.3+)
- `is_sm100a_supported()` - Blackwell (requires CUDA 12.8+)
- `is_sm100f_supported()` - Blackwell feature-set (requires CUDA 12.9+)
- `is_sm110a_supported()`, `is_sm120a_supported()`, `is_sm120f_supported()`, `is_sm121a_supported()`

**APIs decorated with `@backend_requirement`** also provide:
- `api_name.is_compute_capability_supported(cc)` - e.g., `mm_fp4.is_compute_capability_supported(100)`
- `api_name.is_backend_supported("backend")` - e.g., `mm_fp4.is_backend_supported("cudnn")`
- `api_name.is_backend_supported("backend", cc)` - e.g., `mm_fp4.is_backend_supported("cudnn", 80)`

**Example:**
```python
from flashinfer.utils import is_sm90a_supported

def test_hopper_attention():
    if not is_sm90a_supported(torch.device("cuda")):
        pytest.skip("Requires SM90a")
    # Test code...
```

**Common requirements:**

| Feature | Min SM | Check Function |
|---------|--------|----------------|
| FlashAttention-3 | SM90a | `is_sm90a_supported()` |
| MLA Attention | SM100a | `is_sm100a_supported()` |
| FP8 GEMM | SM89+ | `get_compute_capability()[0] >= 9` |

**Note:** `tests/conftest.py` auto-skips tests that trigger OOM, but tests should be written to avoid OOM by using appropriate problem sizes.

## Benchmarking

For SM90 NVFP4 fused/split and MXFP4 fused MegaMoE, use
[`benchmarks/bench_megamoe_fp4_sm90.py`](benchmarks/bench_megamoe_fp4_sm90.py) for timing
and [`tests/moe/test_megamoe_fp4_sm90.py`](tests/moe/test_megamoe_fp4_sm90.py) for correctness checks.

FlashInfer provides a unified benchmarking framework in `benchmarks/flashinfer_benchmark.py`.

**Key features:**
- Supports attention, GEMM, and MOE kernels
- Multiple backends: FlashAttention2/3, cuDNN, CUTLASS, TensorRT-LLM, cuBLAS
- **CUPTI timing (recommended)**: Hardware-level profiling for accurate GPU kernel time
  - Automatically falls back to CUDA events if CUPTI unavailable
  - Install: `pip install -U cupti-python` (requires CUDA 13+)
- Batch testing, reference checking, CSV output

**Quick example:**
```bash
python benchmarks/flashinfer_benchmark.py \
    --routine BatchDecodeWithPagedKVCacheWrapper \
    --backends fa2 cudnn \
    --batch_size 32 --s_kv 2048 \
    --num_qo_heads 32 --num_kv_heads 8 \
    --head_dim_qk 128 --head_dim_vo 128 \
    --page_size 16 --refcheck -vv
```

**Python API:**
```python
from flashinfer.testing import bench_gpu_time

# CUPTI preferred, auto-fallback to CUDA events.
# Returns per-iteration times in milliseconds (a list).
times = bench_gpu_time(
    my_kernel, input_args=(x, y), enable_cupti=True,
    dry_run_iters=10, repeat_iters=30,
)
median_time_ms = statistics.median(times)
```

→ **For complete benchmarking guide, see [`.claude/skills/benchmark-kernel/SKILL.md`](.claude/skills/benchmark-kernel/SKILL.md)**

`flashinfer.gemm.group_gemm_fp8_nt_groupwise_contiguous` is the standalone
CuTe-DSL grouped FP8 API; `group_deepgemm_fp8_nt_groupwise` keeps its original
DeepGEMM-only signature. Both use per-row expert indices, unlike the indptr-based
`group_gemm_fp8_nt_groupwise` API.

For contiguous grouped FP8 GEMM, `python benchmarks/bench_grouped_fp8.py
--production-shapes` compares DeepGEMM and CuTe-DSL with preallocated outputs;
`--cache-probe` measures first-call and new-token-count compilation separately.
Compiled grouped CuTe-DSL kernels are reused across token counts with the same
128-row alignment class, with at most two variants per device and weight shape.
Concurrent misses share one compilation; warm hits do not acquire the compile lock.
The standalone API's optional `validate_indices=True` checks expert-index values
and synchronizes with the CPU. Validate new routing data before CUDA graph
capture; the default path assumes valid indices and performs metadata checks only.

Prims-TS decode keeps raw storage dtypes (`q_dtype`, `k_dtype`, `v_dtype`)
separate from effective MMA dtypes (`qk_dtype`, `pv_dtype`). Transformed
FP8/NVFP4 K/V uses Q's compute dtype; BF16 Q/K with FP8 V uses FP8 PV without
that transform. Kernel predicates, P packing, and attention-sink scaling must
follow effective compute precision; TMA descriptors and raw allocations must
follow storage precision. The public plan accepts separate `k_data_type` and
`v_data_type`, with `kv_data_type` retained as a common-KV compatibility alias.

## Code Linting

Run all pre-commit hooks:

```bash
pre-commit run -a
```

Install hooks to run on every commit:

```bash
pre-commit install
```

## Opening a Pull Request

Follow [`CONTRIBUTING.md`](CONTRIBUTING.md) when opening a PR. The two requirements most often
missed:

- **Keep the default PR template.** Fill in `.github/pull_request_template.md` (Description,
  Related Issues, Checklist, Tests, Reviewer Notes). Do **not** overwrite or replace it with a
  custom or tool-generated description format. The title and description normally become the
  commit title and message on squash-merge, and are what a `git bisect` surfaces months later
  when someone is hunting the owner of a regression.
- **Report before/after numbers for performance work.** A perf PR must include measurements from
  a reproducible benchmark (e.g. `benchmarks/flashinfer_benchmark.py`), naming the GPU and the
  problem sizes. Speedup ratios without absolute numbers, or numbers without a named GPU, are
  not enough.

When preparing changes for a PR, a self-review is strongly recommended: follow
[`.claude/skills/self-review/SKILL.md`](.claude/skills/self-review/SKILL.md), which walks the
diff through the repository's review guidance and PR rules. It is informal and not required —
its purpose is to make review easier for reviewers and contributors alike by catching common
issues before a human looks. Changes made only to test something locally, with no PR intended,
do not need one.

→ **For the complete contribution rules, see [`CONTRIBUTING.md`](CONTRIBUTING.md)**

## Code Review

When reviewing a diff (as an agent or a human), follow the shared focus areas, kernel-review
policy, and effort calibration in [`docs/code_review_guidance.md`](docs/code_review_guidance.md).
Note: unlike human review, agents keep **kernel implementation details in scope** — read the
kernel logic and report bugs, labeling findings by confidence.

→ **For the complete review rule set, see [`docs/code_review_guidance.md`](docs/code_review_guidance.md)**

## Architecture: JIT Compilation System

FlashInfer's JIT system has three layers:

### Layer 1: JitSpec (flashinfer/jit/core.py)

`JitSpec` is an abstract base class defining the kernel-module lifecycle
(`try_load()` / `build()` / `load()`, with the shared `build_and_load()`
template method handling caching, locking, and `FLASHINFER_DISABLE_JIT`).
One subclass per compilation toolchain:

- `JitSpecNvcc` (nvcc/ninja modules, returned by `gen_jit_spec()`) defines:
  - `name`: Unique identifier (URI hash from parameters)
  - `sources`: List of .cu/.cpp files to compile
  - `extra_cuda_cflags`, `extra_cflags`, `extra_ldflags`: Compiler flags
- `JitSpecCuteDsl` (flashinfer/jit/cute_dsl_core.py) caches CuTe-DSL kernels
  (see "CuTe-DSL kernels" under Module Caching below)
- The optional TIRx KDA backend uses a private `JitSpec` in
  `flashinfer/kda_kernels/tirx/cache.py` to cache TVM modules, with source/compiler
  fingerprints and artifact checksums. Its TVM/TIRx dependencies are loaded only
  by explicit `recurrent_kda(..., backend="tirx")`; see `docs/api/kda.rst`.

### JIT Directory Rules

**NEVER write to package directories** - they may be read-only after installation.

| Directory | Writable | Use for |
|-----------|----------|---------|
| `FLASHINFER_GEN_SRC_DIR` | ✓ Yes | Generated source files (Jinja output, copied .cu files) |
| `FLASHINFER_JIT_DIR` | ✓ Yes | Compiled `.so` outputs |
| `FLASHINFER_CSRC_DIR` | ✗ No | Read-only source templates |
| `FLASHINFER_AOT_DIR` | ✗ No | Read-only pre-compiled binaries |

### Compilation Context: Architecture-Specific Compilation

FlashInfer uses `CompilationContext` to manage CUDA architecture targets. Some kernels only work on specific GPU architectures (e.g., Hopper SM90, Blackwell SM100/SM12x).

**How it works:**
- Auto-detects GPUs in system or reads `FLASHINFER_CUDA_ARCH_LIST` environment variable
- CuTe-DSL FP4 MM compilation parallelism can be controlled with `FLASHINFER_MM_FP4_CUTE_DSL_COMPILE_WORKERS`
- JIT modules specify `supported_major_versions=[9, 10, 11, 12]` to limit compilation to specific SM versions
- If GPU not supported → `RuntimeError: No supported CUDA architectures found`

→ **See [`.claude/skills/add-cuda-kernel/SKILL.md`](.claude/skills/add-cuda-kernel/SKILL.md) for usage examples**

### Layer 2: Code Generation

Every `gen_*_module()` function in `flashinfer/jit/` follows this pattern:

```python
def gen_some_module(dtype_in, dtype_out, ...):
    # 1. Compute unique identifier from parameters
    uri = get_some_uri(dtype_in, dtype_out, ...)

    # 2. Create generation directory
    gen_directory = jit_env.FLASHINFER_GEN_SRC_DIR / uri

    # 3. (Optional) Render Jinja template to generate type-specialized config
    # Skip this step if you don't need type specialization
    with open(jit_env.FLASHINFER_CSRC_DIR / "some_customize_config.jinja") as f:
        template = jinja2.Template(f.read())
    config_content = template.render(
        dtype_in=dtype_map[dtype_in],
        dtype_out=dtype_map[dtype_out],
        # ... more parameters
    )
    write_if_different(gen_directory / "some_config.inc", config_content)

    # 4. Copy source files to gen directory
    sources = []
    for fname in ["some_kernel.cu", "some_jit_binding.cu"]:
        shutil.copy(jit_env.FLASHINFER_CSRC_DIR / fname, gen_directory / fname)
        sources.append(gen_directory / fname)

    # 5. Return JitSpec
    return gen_jit_spec(uri, sources, extra_cuda_cflags=[...])
```

**Note**: If your operation doesn't need type specialization, you can skip step 3 entirely and just copy the source files directly.

### Layer 3: Compilation and Loading

`JitSpecNvcc` methods:

- `write_ninja()` - Generates `build.ninja` file
- `build()` - Executes `ninja` to compile sources
- `build_and_load()` - Compiles and loads via TVM-FFI

The generated `build.ninja` file uses nvcc to compile .cu → .cuda.o → .so, then loads via TVM-FFI.

### Jinja Templates (Optional)

**Note: Jinja templates are NOT required.** You can write C++ code directly without templating.

For operations that need type specialization, templates in `csrc/*.jinja` can generate C++ code:

```jinja
// Input template
using DTypeIn = {{ dtype_in }};
using DTypeOut = {{ dtype_out }};
constexpr int PARAM = {{ param_value }};

// After render
using DTypeIn = float16;
using DTypeOut = float16;
constexpr int PARAM = 128;
```

This allows the same CUDA template code to be compiled with different concrete types. However, if your operation doesn't need this, you can skip Jinja and write the `.cu` files directly.

## Directory Structure

```
flashinfer/
├── include/flashinfer/           # Header-only CUDA kernel templates
│   ├── attention/                # Attention kernels
│   ├── gemm/                     # GEMM kernels
│   ├── comm/                     # Communication kernels
│   ├── mma.cuh                   # Matrix multiply utilities
│   ├── utils.cuh                 # Common utilities
│   └── [...]
│
├── csrc/                          # Framework bindings (via TVM-FFI)
│   ├── *.cu                       # Kernel launcher implementations
│   ├── *_jit_binding.cu           # TVM-FFI exports
│   ├── *_customize_config.jinja   # Type config templates (optional)
│   └── [...]
│
├── flashinfer/                    # Python package
│   ├── jit/
│   │   ├── core.py                # JitSpec, compilation infrastructure
│   │   ├── cpp_ext.py             # Ninja build generation
│   │   ├── env.py                 # Workspace paths
│   │   ├── attention/             # Attention module generators
│   │   ├── gemm/                  # GEMM module generators
│   │   ├── fused_moe/             # MOE module generators
│   │   └── [...]
│   ├── gemm/                      # GEMM Python APIs
│   ├── fused_moe/                 # MOE Python APIs
│   ├── comm/                      # Communication Python APIs
│   ├── *.py                       # Other high-level Python APIs
│   ├── aot.py                     # AOT compilation for pre-built packages
│   └── [...]
│
├── tests/                         # Test suite
│   ├── attention/                 # Attention kernel tests
│   ├── gemm/                      # GEMM kernel tests
│   ├── moe/                       # MOE kernel tests
│   ├── comm/                      # Communication tests
│   ├── utils/                     # Utility tests
│   └── conftest.py                # Pytest configuration
│
└── build_backend.py               # PEP 517 build backend
```

### Critical Rule: Framework Separation

**Torch headers MUST NOT be included in `include/` directory files.**

- `include/`: Framework-agnostic CUDA kernels (accept raw pointers)
- `csrc/`: Framework bindings via TVM-FFI (currently PyTorch, but can support other frameworks)

## Adding a New Operation

→ **For complete step-by-step tutorial, see [`.claude/skills/add-cuda-kernel/SKILL.md`](.claude/skills/add-cuda-kernel/SKILL.md)**

**Quick overview of the process:**
1. Write kernel in `include/flashinfer/new_op.cuh` (framework-agnostic, raw pointers)
2. Write launcher in `csrc/new_op.cu` (PyTorch tensor handling)
3. Create TVM-FFI bindings in `csrc/new_op_jit_binding.cu`
4. (Optional) Create Jinja template for type specialization
5. Write JIT module generator in `flashinfer/jit/new_op.py`
6. Write Python API in `flashinfer/new_op.py` with `@functools.cache`
7. Write tests in `tests/`
8. Register in `flashinfer/aot.py` for AOT compilation
9. Export in `flashinfer/__init__.py`
10. Add a `TraceTemplate` in `flashinfer/trace/templates/` and wire it via `@flashinfer_api(trace=...)` (see below)
11. Add an example call in `tests/trace/example.py`, re-run to regenerate `fi_trace_out/`, and commit the new JSON files

### Trace Template Checklist (for new or updated APIs)

Every public API decorated with `@flashinfer_api` should also carry a `trace=` argument so that `fi_trace()` works and auto-dump produces a benchmark definition JSON.

1. **Create or update a `TraceTemplate`** in `flashinfer/trace/templates/<category>.py` (e.g., `norm.py`, `activation.py`, `cascade.py`, `gdn.py`). Define `axes`, `inputs`, `outputs`, and optionally a `reference` function.
2. **Wire the template** to the API: `@flashinfer_api(trace=my_trace)` on the Python function (or class method's `run()`).
3. **Add an example call** in `tests/trace/example.py` that exercises the new trace with realistic shapes.
4. **Regenerate examples**: `rm -rf tests/trace/fi_trace_out && python tests/trace/example.py` — verify the expected JSON appears.
5. **Update the docstring** in `tests/trace/example.py` to list the new file(s).
6. **Run tests**: `pytest tests/trace/ -v` — all template-consistency and end-to-end tests must pass.
7. **Commit the new JSON files** under `tests/trace/fi_trace_out/` alongside the code changes.

**Example implementations:**
- **Simple**: `flashinfer/norm/__init__.py` (RMSNorm) - no Jinja, good starting point
- **Moderate**: `flashinfer/sampling.py` - with Jinja templating
- **Complex**: `flashinfer/decode.py` - plan-run pattern, advanced workspace

## Experimental Code

Experimental backends live under `flashinfer/experimental/`; experimental
public APIs live in core marked with `@flashinfer_experimental_api` (defined in
`flashinfer/api_logging.py`). Using them is an explicit opt-in: calling an
experimental API, or naming an experimental backend with `backend="<name>"`,
needs no environment variable (both warn once). Only automatic selection is
gated: `backend="auto"` skips backends marked `@experimental_backend` unless
`FLASHINFER_ALLOW_EXPERIMENTAL_AUTO_BACKENDS=1`. Experimental code is JIT-only
(never registered in `flashinfer/aot.py`) and tested under
`tests/experimental/`.

→ **When working under `flashinfer/experimental/`, follow
[`flashinfer/experimental/CLAUDE.md`](flashinfer/experimental/CLAUDE.md); the
full policy is [`flashinfer/experimental/README.md`](flashinfer/experimental/README.md).**

## Key Architectural Patterns

### Module Caching

FlashInfer uses two-level caching to avoid recompilation:

1. **Python-level** (`@functools.cache`): In-memory cache of loaded modules
2. **File-level** (`~/.cache/flashinfer/`): Compiled `.so` files on disk

**Cache invalidation** (automatic):
- Source file changes (SHA256 hash)
- Compilation flags change
- CUDA architecture change
- FlashInfer version change

URI computed as: `hash(operation_type + parameters + source_hashes + flags + cuda_arch)`

**Cache management:**
- Clear cache: `rm -rf ~/.cache/flashinfer/`
- Override location: `export FLASHINFER_WORKSPACE_BASE="/scratch"`

**CuTe-DSL kernels** (`flashinfer/jit/cute_dsl_core.py`): `cute.compile()` has no
persistent cache, so `JitSpecCuteDsl` (a `JitSpec` subclass, wrapped by the
`build_and_load_cute_dsl_kernel()` helper) exports compiled kernels as
object files (`export_to_c()`) and reloads them with `cute.runtime.load_module(...,
enable_tvm_ffi=True)`. Artifacts live in `cached_ops/` next to the nvcc modules, one
directory per op family — `cached_ops/<module>_<arch>_cute_dsl/` (e.g.
`nvfp4_quantize_sm100a_cute_dsl/`) holding one `meta.json` plus one `.o` per
specialization. The arch comes from the DSL's compile target (`CUTE_DSL_ARCH` or the
current device) since the artifacts are single-arch, unlike nvcc fatbins.
Invalidation is module-granular: changed DSL/Torch/TVM-FFI distributions, Python
ABI, or
kernel-source SHA256 wipes and lazily rebuilds the module. Reference usage:
`flashinfer/quantization/kernels/nvfp4_quantize.py`. Disable with
`FLASHINFER_CUTE_DSL_DISABLE_CACHE=1`. The SHA256 covers only the files a caller
passes as `extra_key_files` (often just `__file__`); with
`FLASHINFER_CUTE_DSL_STRICT_CACHE_KEY=1` it also covers every flashinfer module
those files can run. Source analysis and hashing live in
`flashinfer/jit/_cute_dsl_source.py`; `cute_dsl_core.py` manages compilation and
cached artifacts. Plain modules are keyed by content with everything they
import, including literal `importlib.import_module` calls. Package initializers
are resolved name by name: a name looked up in a package follows the binding
that defines it (a re-export to its source, a function into its body, a module
through the attributes read from it) plus the submodule of that name. The code
an initializer runs at import, including evaluated function annotations, is
followed for every keyed file in its package. Star imports remain covered when
they can replace an earlier binding; lazy fallbacks are followed when a binding
may be absent.
Each initializer is keyed by that code and the import lines of used names, so a
new re-export or `__all__` entry invalidates nothing. Not keyed: import-time
side effects of modules that keyed code neither imports nor uses, nonliteral
dynamic imports, and runtime file reads; list the latter in `extra_key_files`.
`version.py` and `_build_meta.py` are excluded to avoid keying on the checkout
commit. This is intended for long-lived caches shared across commits.
Task-scheduled attention shares this cache through
`flashinfer/attention/prims_ts/_compile_cache.py`; its keys cover compile specs,
options, and embedded device-capacity constants. It hashes the attention source
tree, including shared preparation, masks, and reductions. Occupancy queries
are cached per device and cluster size by `_hardware.py`.

### Dispatch Macros

Handle combinatorial parameter spaces:

```cpp
DISPATCH_DTYPE(input_dtype, DTypeIn, {
  DISPATCH_DTYPE(output_dtype, DTypeOut, {
    DISPATCH_BLOCK_SIZE(block_size, BLOCK_SIZE, {
      LaunchKernel<DTypeIn, DTypeOut, BLOCK_SIZE>(...);
    });
  });
});
```

Defined in `.jinja` files and expanded after rendering.

## API Logging with @flashinfer_api

FlashInfer provides the `@flashinfer_api` decorator for debugging API calls.

**Key features:**
- **Crash-safe**: Logs inputs BEFORE execution (preserves info even if kernel crashes)
- **Zero overhead when disabled**: `FLASHINFER_LOGLEVEL=0` (default)
- **Multiple verbosity levels**: 0 (off), 1 (names), 3 (inputs/outputs), 5 (+ statistics)
- **CUDA graph compatible**: Auto-skips stats during graph capture

**Quick usage:**
```bash
# Enable detailed logging
export FLASHINFER_LOGLEVEL=3              # 0, 1, 3, or 5
export FLASHINFER_LOGDEST=debug.log       # stdout, stderr, or file path

python my_script.py
```

**Why use this?**
- Debug CUDA crashes (see inputs that caused crash)
- Track tensor shapes/dtypes through pipeline
- Detect NaN/Inf issues (level 5)

→ **For complete debugging guide, see [`.claude/skills/debug-cuda-crash/SKILL.md`](.claude/skills/debug-cuda-crash/SKILL.md)**

## Debugging

### Enable Logging

```bash
export FLASHINFER_JIT_VERBOSE=1      # Verbose JIT output
export FLASHINFER_JIT_DEBUG=1        # Debug symbols, -O0
export FLASHINFER_LOGLEVEL=3         # API logging (0=off, 1=basic, 3=detailed)
export FLASHINFER_LOGDEST=stdout
```

### Inspect Generated Code

```bash
# Generated sources
ls -la ~/.cache/flashinfer/0.6.0/*/generated/

# Compiled modules
ls -la ~/.cache/flashinfer/0.6.0/*/cached_ops/

# Build files
cat ~/.cache/flashinfer/0.6.0/*/cached_ops/*/build.ninja
```

### Environment Variables

```bash
# Compilation
export FLASHINFER_NVCC_THREADS=4              # Threads per nvcc process (--threads=N)
export MAX_JOBS=4                             # Parallel ninja jobs (nvcc processes)
export FLASHINFER_CUDA_ARCH_LIST="8.0 9.0a"  # Target architectures
# Memory note: total compilation memory ≈ MAX_JOBS × FLASHINFER_NVCC_THREADS × per-thread mem.

# Behavior
export FLASHINFER_WORKSPACE_BASE="/scratch"   # Custom cache directory
```

#### Full Environment Variable Reference

Already covered above (Quick Reference / Debugging): `FLASHINFER_LOGLEVEL`,
`FLASHINFER_LOGDEST`, `FLASHINFER_JIT_VERBOSE`, `FLASHINFER_JIT_DEBUG`,
`FLASHINFER_CUDA_ARCH_LIST`, `FLASHINFER_NVCC_THREADS`,
`FLASHINFER_WORKSPACE_BASE`, `MAX_JOBS`.

The remaining `FLASHINFER_*` knobs read by the code are listed below. Defaults
match what the code uses today; values are strings unless noted.

##### JIT / Build Toolchain

| Variable | Default | Read in | Effect |
|----------|---------|---------|--------|
| `FLASHINFER_DISABLE_JIT` | unset | `flashinfer/jit/core.py` | If set (any non-empty value), JIT compilation is refused and modules must already exist in the cache or be provided via AOT packages. |
| `FLASHINFER_CUTE_DSL_DISABLE_CACHE` | `0` | `flashinfer/jit/cute_dsl_core.py` | `1` disables the on-disk cache for JIT-compiled CuTe-DSL kernels (every process recompiles via `cute.compile`). |
| `FLASHINFER_CUTE_DSL_STRICT_CACHE_KEY` | `0` | `flashinfer/jit/cute_dsl_core.py` | `1` extends each CuTe-DSL disk-cache key from the caller's `extra_key_files` to their in-package import closure, so editing an imported helper invalidates the kernel. Changes every key, so the first run after enabling recompiles. Set by CI's persistent compile cache. |
| `FLASHINFER_JIT_BUILD_LOG` | unset | `flashinfer/jit/core.py` | Path of a JSONL file that gets one record (`toolchain`, `name`, `seconds`) per build made through `JitSpec.build_and_load()` or a bulk `build_jit_specs()` prebuild: nvcc modules (ninja no-ops included), CuTe-DSL kernels and TIRx modules. Triton, cuTile and the cake nvcc/NVRTC side paths are not recorded. Summarize with `scripts/summarize_jit_builds.py`. |
| `FLASHINFER_CUDNN_FROST_PTXAS` | unset | `flashinfer/fused_moe/backends/cudnn_frost/compiler.py` | Select an external PTXAS executable path for all Frost dtypes; unset or empty uses the bundled assembler. The executable path, version, and SHA-256 participate in compiler and tactic cache identities. MoE engines snapshot this configuration at build time; call `layer.refresh_backends()` or reconstruct the layer to adopt changes. |
| `FLASHINFER_DISABLE_VERSION_CHECK` | unset | `flashinfer/jit/env.py` | Skip the check that pins flashinfer-cubin and flashinfer-jit-cache to the installed flashinfer-python. Without it, mismatched kernel wheels (typically left behind by an upgrade) are ignored with a warning at import, falling back to on-demand cubin downloads and JIT compilation until `flashinfer download-kernels` installs matching ones. Set it only when you intentionally mix versions. |
| `FLASHINFER_JIT_LINEINFO` | `0` | `flashinfer/jit/core.py` | `1` adds `-lineinfo` to nvcc so profiler / `cuda-gdb` can map PTX back to CUDA source. |
| `FLASHINFER_JIT_PREBUILD_MAX_JOBS` | unset (uses `MAX_JOBS`) | `flashinfer/jit/core.py` | Maximum parallel Ninja jobs for a bulk `build_jit_specs()` prebuild. This does not control ordinary on-demand module builds. The sharded test runner sets it to the host-wide automatic build budget before reducing each worker's `MAX_JOBS`; an explicit value is preserved. |
| `FLASHINFER_NVCC` | `$cuda_home/bin/nvcc` | `flashinfer/jit/cpp_ext.py` | Override the nvcc binary used by the JIT (useful for sccache wrappers or non-default CUDA installs). |
| `FLASHINFER_NVCC_LAUNCHER` | `""` (`sccache` in the devcontainers) | `flashinfer/jit/cpp_ext.py` | Optional launcher prefix for nvcc (e.g. `ccache`, `sccache`). Combined with `FLASHINFER_NVCC`. |
| `FLASHINFER_CXX_LAUNCHER` | `""` (`sccache` in the devcontainers) | `flashinfer/jit/cpp_ext.py` | Same idea as `FLASHINFER_NVCC_LAUNCHER` but for the host C++ compiler. |
| `FLASHINFER_FMHA_V2_VERBOSE` | unset | `flashinfer/jit/attention/fmha_v2/fmha_library.py` (FMHA v2 codegen, C++ side) | When set, the FMHA-v2 codegen / runtime prints verbose dispatcher diagnostics. Leave unset for normal runs. |
| `FLASHINFER_EXTRA_CFLAGS` | unset | `flashinfer/jit/cpp_ext.py` | Extra compiler flags passed to the host C++ compiler. |
| `FLASHINFER_EXTRA_CUDAFLAGS` | unset | `flashinfer/jit/cpp_ext.py` | Extra compiler flags passed to `nvcc`. |
| `FLASHINFER_EXTRA_LDFLAGS` | unset | `flashinfer/jit/cpp_ext.py` | Extra linker flags passed to the linker. |
| `FLASHINFER_CAKE_GDN_VALIDATE_SLOTS` | `0` | `flashinfer/gdn_decode.py` | `1` enables extra slot-validation checks in the CAKE GDN path for debugging or invariant validation. Leave disabled in normal runs to avoid extra validation overhead. |
| `FLASHINFER_CAKE_DSA_CHECK_DST_MAP` | `0` | `flashinfer/dsa_sparse_attention.py` | Fallback of the `check_dst_map` keyword of `dsa_sparse_attention` / `dsa_sparse_attention_varlen` when it is `None`: unset, empty or `0` leaves the values of a caller-provided `dkv_dst_map` unchecked; any other value makes the backward validate them (every destination row in `[0, S_dst)` of `dkv_acc`) at the cost of one device synchronization and raise `ValueError` on a violation. The kernel itself does not range-check the map; the package backend reads no environment variable. |
| `FLASHINFER_CAKE_DSA_DKV_DIRECT` | `auto` | `flashinfer/dsa_sparse_attention.py` | Fallback of the `dkv_direct` keyword of the two public entries when it is `None`, read on every call: `auto` -> `None` (the record's size rule: at least `dkv_direct.min_keys_per_query` keys per query token), `1` -> `True` (direct dK / dV accumulation whenever the natural-layout stages are registered), `0` -> `False` (the accumulate-and-cast path); any other value raises `ValueError`. The resolved value is part of the backend's binding key. |
| `FLASHINFER_CAKE_DSA_TRAIN_BINDING_CACHE_CAPACITY` | `64` | `flashinfer/dsa_sparse_attention.py` | Capacity of the CAKE DSA training binding cache (`cake_backend.BINDING_CACHE`, least recently used first out): read once when `flashinfer.dsa_sparse_attention` is imported (a value that is not a positive integer raises `ValueError`; unset or blank keeps the default) and applied through `BINDING_CACHE.configure(capacity=...)` when the wrapper first reaches the backend. Direct users of the package call `configure` themselves. |
| `FLASHINFER_CAKE_LM_HEAD_LOSS_GEMM_TUNING` | unset | `flashinfer/chunked_lm_head.py` | Fallback of the `gemm_tuning` keyword of `chunked_lm_head_loss` / `chunked_lm_head_logprob` when it is `None`: a JSON object of explicit per-GEMM knobs, e.g. `{"dx": {"tile_n": 256, "k_slices": 2}}` (the package backend takes the same dict as a keyword and reads no environment variable). Unset / empty: the per-geometry rules choose. Invalid JSON or a non-object is a `ValueError`. |
| `FLASHINFER_CAKE_LM_HEAD_LOSS_FUSE_DW_CAST` | `1` | `flashinfer/chunked_lm_head.py` | Fallback of the `fuse_dw_cast` keyword when it is `None`: `0` turns off the fused weight-gradient cast (every chunk then accumulates into the FP32 `dW_acc` and a separate kernel applies the upstream scale and the output cast). `dW` is bitwise the same either way. |
| `FLASHINFER_CAKE_LM_HEAD_LOSS_DX_FINALIZE` | `1` | `flashinfer/chunked_lm_head.py` | Fallback of the `dx_finalize` keyword when it is `None`: `0` turns off the fused dX finalize (a K-sliced dX GEMM's FP32 slabs are then added by in-place `add_` launches and a compacted call's `dX` is cast, zero-filled and `index_copy_`-scattered on the host instead of by the `slab_sum` and `scale_cast_scatter_bf16` kernels). Bitwise the same `dX`; resolved once per forward. |
| `FLASHINFER_CAKE_LM_HEAD_LOSS_DW_STREAM` | `auto` | `flashinfer/chunked_lm_head.py` | Fallback of the `dw_stream` keyword when it is `None` (`0` / `false` / `off` / `no` -> never, `1` / `true` / `on` / `yes` -> every multi-chunk call, anything else -> `auto` = calls of three or more chunks): launches each chunk's weight-gradient accumulate GEMM on a per-device side stream (forked after the chunk's row gradients, joined before the chunk buffer is reused and before the call returns) so it overlaps the tail of the chunk's dX GEMM. Same kernels in the same per-kernel order; bitwise the same outputs. |
| `FLASHINFER_CAKE_LM_HEAD_LOSS_HIDDEN_COUNT` | `1` | `flashinfer/chunked_lm_head.py` | Fallback of the `hidden_count` keyword when it is `None`: `0` turns off the hidden valid-row count of the eager entry points (served to calls of three or more chunks): a compacted call then counts its valid rows and forms their index on the host before its first launch instead of on the device with chunk 0's gather kernel and device-count logits GEMM queued ahead of the read-back. Bitwise the same outputs. |
| `FLASHINFER_CAKE_LM_HEAD_LOSS_COMPACT_ROWS` | `1` | `flashinfer/chunked_lm_head.py` | Fallback of the `compact_rows` keyword when it is `None`: `0` turns off valid-row compaction (the chunk loop then covers every row instead of only the rows whose label is not `-100`). |
| `FLASHINFER_CAKE_LM_HEAD_LOSS_BINDING_CACHE` | `1` | `flashinfer/chunked_lm_head.py` | Read once when `flashinfer.chunked_lm_head` is imported and applied through `cake_backend.BINDING_CACHE.configure` when the wrapper first reaches the backend (the first call of either entry point): `0` disables the per-process LRU cache of compiled bindings. Direct users of the package call `configure` themselves. |
| `FLASHINFER_CAKE_LM_HEAD_LOSS_BINDING_CACHE_CAPACITY` | `64` | `flashinfer/chunked_lm_head.py` | Capacity of that binding cache, read once at the same import (a value that is not a positive integer raises `ValueError`; unset or blank keeps the default) and applied through `configure` on the same first use. |
| `FLASHINFER_JIT_CACHE_PROVIDER_ARCHS` | required | `flashinfer-jit-cache/build_backend.py` | Space-separated provider architectures added to a shim wheel's exact `Requires-Dist` metadata. |
| `FLASHINFER_JIT_CACHE_PROVIDER_ARCH` | required | `flashinfer-jit-cache-provider/package_config.py` | Select exactly one architecture, such as `9.0a` or `sm120f`, when building a binary provider wheel. |

Release/nightly JIT-cache provider builds run sccache in server-side mode and
wrap the complete wheel build with a no-output watchdog. After 90 minutes with
no build output, the watchdog captures system, process, compiler, and sccache
state; terminates the wrapped build process group (TERM, then KILL after a
two-minute grace period); and exits with status 124. Compiler discovery scans
the whole build container, so diagnostics include compiler processes owned by
the sccache daemon even though they are outside the wrapped process group. The
CI `docker run --rm` boundary provides final cleanup for those processes.

##### Cubin / Artifact Loader

| Variable | Default | Read in | Effect |
|----------|---------|---------|--------|
| `FLASHINFER_CUBIN_DIR` | `<FLASHINFER_WORKSPACE_BASE>/.cache/flashinfer/cubins` | `flashinfer/jit/env.py` (exposed via `flashinfer/__main__.py show-config`) | Local directory used to cache downloaded cubins. Override to share a cache between users. |
| `FLASHINFER_CUBINS_REPOSITORY` | `https://edge.urm.nvidia.com/artifactory/sw-kernelinferencelibrary-public-generic-local` | `flashinfer/jit/cubin_loader.py` | Base URL the loader downloads cubins from. Point to a mirror for offline or air-gapped setups. |
| `FLASHINFER_CUBIN_CHECKSUM_DISABLED` | unset | `flashinfer/jit/cubin_loader.py` | If set, skip SHA checksum verification of downloaded cubins. Debug aid only. |
| `FLASHINFER_CUBIN_DOWNLOAD_THREADS` | `4` | `flashinfer/artifacts.py` | Thread-pool size used by `flashinfer artifacts download`. |
| `FLASHINFER_CUBIN_RETRY_WINDOW_SECONDS` | `0` | `flashinfer/artifacts.py` | Total wall-clock retry window for `download_artifacts()`. A failed artifact keeps re-attempting, with capped exponential backoff, until the window closes (e.g. `86400` for a 24-hour nightly window). `0` keeps only `download_file()`'s own per-call retry budget. |
| `FLASHINFER_NO_DOWNLOAD` | unset | `flashinfer/jit/cubin_loader.py` | Hard-fail if a cubin is missing locally instead of attempting to download. Useful in CI / locked-down environments. |
| `FLASHINFER_DSL_FMHA_LOCAL_DIR` | unset | `flashinfer/attention/cute_dsl/fmha.py` | Path to a local checkout of the CuTe-DSL FMHA kernel sources. The loader checks here before downloading. |
| `FLASHINFER_LOGGING_LEVEL` | `INFO` | `flashinfer/artifacts.py`, `flashinfer/jit/core.py` | Python logging level for the artifacts/cubin loader and the JIT compiler (`DEBUG`/`INFO`/`WARNING`/`ERROR`). Distinct from `FLASHINFER_LOGLEVEL`. |
| `FLASHINFER_DISABLE_TINYGEMM2_SM100` | `0` | `flashinfer/gemm/routergemm.py` | Set to `1` to disable the generated SM100/SM103 `tinygemm2` backend and force the dispatcher to fall back to the legacy implementation. Useful as an escape hatch when debugging backend-selection or kernel issues on Blackwell systems. |

##### API Dump / Logging Extensions

These complement `FLASHINFER_LOGLEVEL`/`FLASHINFER_LOGDEST` and are all read in `flashinfer/api_logging.py`.

| Variable | Default | Effect |
|----------|---------|--------|
| `FLASHINFER_DUMP_DIR` | `flashinfer_dumps` | Directory where `@flashinfer_api` writes per-call tensor dumps when logging is enabled. |
| `FLASHINFER_DUMP_INCLUDE` | `""` | Comma-separated allow-list of API names; only matching calls are dumped. Wildcards (`*`) supported. |
| `FLASHINFER_DUMP_EXCLUDE` | `""` | Comma-separated deny-list of API names; matching calls skip the (expensive) stats / dump path. |
| `FLASHINFER_DUMP_MAX_COUNT` | `1000` | Hard cap on number of dumped events; once reached, further dumps are dropped with a warning. |
| `FLASHINFER_DUMP_MAX_SIZE_GB` | `20` | Hard cap on total dump size in **gigabytes** (GB) written to `FLASHINFER_DUMP_DIR` (parsed as `float`; see `_DUMP_MAX_SIZE_GB` in `flashinfer/api_logging.py`). |
| `FLASHINFER_DUMP_SAFETENSORS` | `0` | `1` writes dumps as `.safetensors` (portable / Hugging Face-style); otherwise plain `.pt`. |

##### Trace Capture

Used by `flashinfer.trace` / `fi_trace`.

| Variable | Default | Read in | Effect |
|----------|---------|---------|--------|
| `FLASHINFER_TRACE_DUMP` | unset | `flashinfer/fi_trace.py` | If set, decorated APIs auto-dump benchmark-definition JSON for each call (the "fi_trace" feature). |
| `FLASHINFER_TRACE_DUMP_DIR` | cwd | `flashinfer/fi_trace.py` | Directory where the trace JSON files are written. |
| `FLASHINFER_TRACE_APPLY` | `0` | `flashinfer/__init__.py`, `flashinfer/trace_apply/config.py` | Set to `1` to enable Trace Apply (runtime kernel substitution) when FlashInfer is imported. |
| `FLASHINFER_TRACE_APPLY_PATH` | unset | `flashinfer/trace_apply/config.py` | Directory from which deployment-configured solutions are loaded for Trace Apply. |

##### Validation / Autotuning / Routing / Kernel Selection

| Variable | Default | Read in | Effect |
|----------|---------|---------|--------|
| `FLASHINFER_ALLOW_EXPERIMENTAL_AUTO_BACKENDS` | unset | `flashinfer/utils.py` (`backend_requirement`), `flashinfer/api_logging.py` | `1` lets `backend="auto"` (dispatch heuristics and autotuning) select backends marked `@experimental_backend` (see the "Experimental Code" section). Explicit use never needs it: calling an `@flashinfer_experimental_api` function or passing `backend="<experimental>"` always works and emits an `ExperimentalWarning` once. Without it, automatic selection considers only stable backends. |
| `FLASHINFER_VALIDATE_INPUTS` | `0` | `flashinfer/mla/_core.py` (MLA wrapper) | Non-zero / non-empty value enables defensive input validation inside the MLA wrapper. Device-synchronizing checks are skipped during CUDA Graph capture. Adds host-side overhead; intended for debugging. |
| `FLASHINFER_AUTOTUNER_LOAD_FROM_FILE` | `0` | `flashinfer/autotuner/autotuner.py` | `1` loads previously serialized autotune results from disk instead of re-running the search. |
| `FLASHINFER_DIST_AWARE_AUTOTUNE` | `0` | `flashinfer/fused_moe/da_config.py` | `1` enables experimental distribution-aware autotune and kernel dispatch (TRT-LLM MoE only). |
| `FLASHINFER_SM107_FUSED_ROUTE_CHECK` | `1` | `flashinfer/moe_ep/backends/mega/kernel/sm107/validation.py` | `0` uses Torch routing checks instead of the fused checker. Both paths check expert IDs, duplicates, and finite scores, including during graph replay. |
| `FLASHINFER_SM107_PREP_TVM_FFI` | `1` | `flashinfer/moe_ep/backends/mega/kernel/sm107/routing_validation.py` | `0` launches the fused routing checker through CuTe's DLPack bindings instead of TVM-FFI. Set before warmup; each mode needs its own compiled kernel. |
| `FLASHINFER_MEGAMOE_LOG_COMPILE_SPEC` | unset | `flashinfer/moe_ep/kernel_src/sm120/split_cutedsl_megakernel/shim/runtime.py` | Exactly `1` prints MegaMoE split-kernel compile choices, workspace memory diagnostics, and frontend compile/capture timings; other values leave these diagnostics off. |
| `FLASHINFER_DA_DISTRIBUTIONS` | built-in distribution catalog | `flashinfer/fused_moe/da_config.py` | Comma-separated training distributions used by the experimental TRT-LLM distribution-aware MoE autotuner. |
| `FLASHINFER_PRIMS_TS_DEBUG_CHECKS` | `0` | `flashinfer/prims_ts/batched_gemm/batched_gemm_kernel.py`, `batched_gemm_run.py`, `flashinfer/prims_ts/gemm/kernel.py` | `1`/`true`/`yes`/`on` enables exhaustive deadlock/race checks and TaskManager verification in Prims-TS batched GEMM and dense GEMM. Production JIT leaves both off. `tests/prims_ts` turns the variable on for `test_batched_gemm_*` and `test_gemm_*`. |
| `FLASHINFER_PRIMS_TS_COMPILE_OPTIONS` | `--opt-level 2` | `flashinfer/prims_ts/utils.py`, `batched_gemm_run.py` | Extra CuTe-DSL / nvc compile options passed when building Prims-TS kernels. |
| `FLASHINFER_PRIMS_TS_BF16_CONSTANT_WEIGHTS` | unset | `flashinfer/prims_ts/batched_gemm/batched_gemm_run.py` | `1` fills BF16 autotune/cast-A weight tensors with constant `1.0` (debug / deterministic fill). |
| `FLASHINFER_PRIMS_TS_CASTA_CONSTANT_ACT` | unset | `flashinfer/prims_ts/batched_gemm/batched_gemm_run.py` | `1` fills cast-A / BF16 activation tensors with constant `1.0` during autotune buffer setup. |
| `FLASHINFER_PRIMS_TS_CASTA_PATTERN_ACT` | unset | `flashinfer/prims_ts/batched_gemm/batched_gemm_run.py` | `1` fills activations with a quarter-wise pattern `(1,2,4,6)` instead of random data. |
| `FLASHINFER_PRIMS_TS_CASTA_PATTERN_WEIGHTS` | unset | `flashinfer/prims_ts/batched_gemm/batched_gemm_run.py` | `1` fills cast-A weights with a deterministic pattern (takes precedence over constant fill). |
| `FLASHINFER_PRIMS_TS_CASTA_CONSTANT_WEIGHTS` | unset | `flashinfer/prims_ts/batched_gemm/batched_gemm_run.py` | `1` fills cast-A weights with constant values when pattern fill is not set. |
| `FLASHINFER_AUTOTUNE_DIR` | unset | `flashinfer/mla/_sparse_mla_sm120/_calibration.py`, `flashinfer/comm/pcie_ipc_tuning.py` | Override the disk path for tuning cache files (sparse-MLA SM120 cpb calibration constants, and the PCIe IPC all-reduce). Falls back to `FLASHINFER_WORKSPACE_DIR` when unset. |
| `FLASHINFER_AUTOTUNE_TIMER` | unset (auto) | `flashinfer/autotuner/autotuner.py` | Selects the autotuner's per-tactic timer: `globaltimer` forces the GPU `%globaltimer` register, `cuda_event` forces `cudaEvent`, unset/anything-else auto-detects (uses `%globaltimer` only when Confidential Computing is detected). Under CC `cudaEventElapsedTime` is unreliable (can go negative), so the globaltimer path keeps tactic ranking stable. |
| `FLASHINFER_CUTILE_AUTOTUNE_DISABLED` | `0` | `flashinfer/quantization/kernels/cutile/rope_quantize_fp8_cutile.py` | Non-zero skips exhaustive cuTile RoPE-FP8 tuning and uses the built-in token-count heuristic. |
| `FLASHINFER_CONFIDENTIAL_COMPUTE` | unset | `flashinfer/utils.py` | Override NVIDIA Confidential Computing (CC) auto-detection used by `is_confidential_compute()` (which drives the autotuner timer above): `1` forces CC, `0` forces non-CC. Useful for CI or hosts without `pynvml`. |
| `FLASHINFER_AUTOTUNE_CACHE_DIR` | `FLASHINFER_CACHE_DIR/autotune` | `flashinfer/autotune_cache.py` | Root **directory** of the managed v2 autotune store used by `autotune_v2()` (placement only; distinct from the MLA-specific `FLASHINFER_AUTOTUNE_DIR` above). |
| `FLASHINFER_RAGGED_AUTO_BACKEND_ORDER` | unset | `flashinfer/prefill.py` | Comma-separated backend order for `BatchPrefillWithRaggedKVCacheWrapper` under `backend="auto"` on Blackwell (e.g. `cutlass,cudnn`, or a single name to pin one). Overrides both the per-shape table and the global default; eligibility is still enforced, so a backend that cannot serve the problem is skipped rather than forced. For benchmarking and regression bisection on hardware whose ranking differs from the measured B200 defaults. |
| `FLASHINFER_NVFP4_QUANTIZE_USE_TMA` | unset | `flashinfer/quantization/kernels/nvfp4_quantize.py` | Overrides the NVFP4 CuTe-DSL TMA-kernel heuristic: `0` forces the vectorized-load kernel, `1` enables the shape heuristic on every architecture and input dtype. Unset, TMA is used only where it was measured to win (SM107 with FP16/BF16 input, layout-specific M/K crossovers). Never taken for fp8 input or an active 4over6 recipe, whatever the setting. For benchmarking and re-tuning on new hardware. |
| `FLASHINFER_NVFP4_PER_TOKEN_THREADS` | shape-dependent (`128`-`512`) | `flashinfer/quantization/kernels/nvfp4_quantize.py` | Override the CTA thread count for per-token NVFP4 quantization. Intended for benchmarking; invalid or unsupported values can make kernel compilation or launch fail. |
| `FLASHINFER_KDA_AFFINE_POLICY` | `model` | `flashinfer/cake_kda_tf32_runtime.py` | Set to `legacy` to restore the pre-2026-09-23 BF16 affine-split gate for A/B measurements. `model`, `packed`, and unrecognized values use the measured cost model. |
| `FLASHINFER_KDA_AFFINE_WINDOW_WAVES` | `auto` | `flashinfer/cake_kda_tf32_runtime.py` | Override resident affine-composite windows in CTA waves. Integer values are clamped to `1`-`3`; unset, `auto`, and invalid values let the cost model choose. |
| `FLASHINFER_KDA_AFFINE_FUSED_EPILOGUE` | `1` | `flashinfer/cake_kda_tf32_runtime.py` | `0` restores the equivalent PyTorch checkpoint/tail/final-state epilogue for A/B testing; any other value keeps the fused epilogue. |
| `FLASHINFER_KDA_PLAN_CACHE_MAX_BYTES` | `8589934592` (8 GiB) | `flashinfer/kda_prefill.py` | Maximum workspace bytes retained by the KDA prefill plan cache. Must parse as a positive integer. |
| `FLASHINFER_KDA_PTXAS` | auto-discovered | `flashinfer/kda_kernels/ptx/static_runtime.py` | Path to a `ptxas` executable at version 13.2 or newer for assembling retained PTX 9.2 kernels. When unset, searches the `nvidia-cuda-nvcc` wheel, `CUDA_HOME`/`CUDA_PATH`, `PATH`, then `/usr/local/cuda/bin/ptxas`. |
| `FLASHINFER_KDA_PTX_CACHE_DIR` | `FLASHINFER_JIT_DIR/kda_ptx` | `flashinfer/kda_kernels/ptx/static_runtime.py` | Override the cache directory for cubins assembled from retained KDA PTX. Defaults to `FLASHINFER_JIT_DIR/kda_ptx`; the selected directory is created on demand. |
| `FLASHINFER_SM90_CAKE_BF16_DEV_LAUNCHER` | unset | `flashinfer/moe_ep/kernel_src/sm90/cake_bf16_megamoe/shim/cake_gemm.py` | Development-only factory for building the SM90 BF16 CAKE grouped-GEMM kernels in process. Set to `module` or `module:attribute` (default attribute: `create_grouped_gemm`); unset, empty, or `0` selects the shipped launcher. |
| `FLASHINFER_SM90_CAKE_BF16_COMBINE_WIRE` | shape-dependent | `flashinfer/moe_ep/kernel_src/sm90/cake_bf16_megamoe/shim/cake_runner.py` | When `combine_wire=None`, select `per_route`, `prereduced`, or `prereduced_hilo`; unset chooses `per_route` for token capacities up to 8 and `prereduced` above. Every rank in a pipe must use the same value. |
| `FLASHINFER_GDN_SM107_PREFILL_KERNEL` | `sm107` | `flashinfer/gdn_kernels/blackwell/gdn_prefill.py` | `sm100` makes `chunk_gated_delta_rule` on SM107 (Rubin) run the Blackwell chunked-prefill kernel class instead of the Rubin subclass. For A/B measurements and bisection. |
| `FLASHINFER_QKV_EXCHANGE_FOLD_BARRIER` | `auto` | `flashinfer/comm/cake_qkv_norm_rope_exchange_overlap.py` | Launch path of the fused QKV exchange session: `1` the producer with the folded cross-rank rendezvous (two launches per call), `0` the separate barrier kernel before the producer (three launches), `auto` the architecture default (folded on SM90 / SM100 / SM103 / SM107). The `fold_barrier` keyword takes precedence. Every rank of a group must use the same value. |
| `FLASHINFER_QKV_EXCHANGE_SHAPE_GATES` | `auto` | `flashinfer/comm/cake_qkv_norm_rope_exchange_overlap.py` | Row-count gates of the session's host plan: `auto` the per-(architecture, world size) default, `off` every gate off, or `gate=value` pairs over `words_capacity_small_m`, `reserved_sms_large_m`, `per_dest_large_m` and `pair_off_large_m` (`0` = never, `auto` = that gate's default, `all` = every row for `words_capacity_small_m`). The first three change the host plan only; `pair_off_large_m` (off everywhere) runs the rows it fires on with the single-CTA producer, which no delivered Blackwell route ships today; SM90 with 8 ranks ships `words_capacity_small_m=all`; no gate changes the stored bits. Every rank of a group must use the same value. |
| `FLASHINFER_QKV_EXCHANGE_PRODUCER_VARIANT` | `auto` | `flashinfer/comm/cake_qkv_norm_rope_exchange_overlap.py` | Producer build of the session: `auto` the per-(architecture, world size) default (`bulk_ow256` on SM100 with 8 ranks, `ow256` on SM100 with 4 ranks and on SM103 with 8 or 4 ranks, `bulk` on SM107 with 4 ranks, the baseline build otherwise), `shipped` the baseline build, or a delivered build name; a build that is not delivered for the session's tile width raises and names the delivered ones. Every build stores identical bits. Every rank of a group must use the same value. |
| `FLASHINFER_QKV_EXCHANGE_GROUP_M_SCALE` | `auto` | `flashinfer/comm/cake_qkv_norm_rope_exchange_overlap.py` | Group scaling of the session's host plan: `auto` the per-(architecture, world size) default (`2` on SM100 with 4 or 2 ranks and on SM103, `1` otherwise), or `1` / `2` / `4`; applies to grouped rows with a hidden size of at least 6144 and changes the tile order only. Every rank of a group must use the same value. |
| `FLASHINFER_TOPK_ALGO` | unset | `flashinfer/topk.py` | Force a specific top-k backend (otherwise the dispatcher chooses based on shape/dtype/mode via benefit gates): `default` (radix), `clusters` (SM100), `cub` (cub::DeviceBatchedTopK; bypasses the benefit gates). Used for benchmarking / regression bisection. |
| `FLASHINFER_USE_CUDA_NORM` | `0` | `flashinfer/norm/__init__.py` | `1` switches the norm path from the default backend to the legacy CUDA-only kernels. Diagnostic toggle. |
| `FLASHINFER_DISABLE_EAGLE_VERIFY_FP8KV_SM90` | `0` | `flashinfer/prefill.py` | `1` keeps `BatchPrefillWithPagedKVCacheWrapper.run` on the FA2 kernel for the speculative-decoding target-verify problems that the specialized SM90 fp8-KV split-KV kernel (`csrc/eagle_verify_fp8kv_sm90.cu`) would otherwise serve: `plan(..., uniform_q_len=4)` with a custom mask, 4 query heads per KV head, head_dim 256, page size 1, NHD, bf16 q/o, e4m3 KV, compute capability 9.0. Read at plan and run time. The `uniform_q_len` plan hint itself is unaffected. |
| `FLASHINFER_SPECIALIZED_KERNEL_DISABLE` | unset | `flashinfer/mla_kv_pack.py` | `1` forces `concat_mla_kv_quant_fp8` onto its composable PyTorch fallback (two fp8 casts + two strided copies, same bytes) instead of the fused MLA K/V pack kernel. Read at call time; the kill switch for A/B testing the fused path. |
| `FLASHINFER_ROUTING_FORCE_BLOCK_PER_TOKEN` | unset | `csrc/fused_moe/trtllm_backend/trtllm_fused_moe_routing_custom.cuh` | Forces the TRT-LLM MoE custom-routing kernel into "one-block-per-token" mode regardless of the active routing policy. Mainly used to reproduce specific perf points. |
| `FLASHINFER_B12X_MICRO_SHARE_INPUT` | `1` | `flashinfer/fused_moe/cute_dsl/blackwell_sm12x/moe_dispatch.py` | `0` disables the B12x MoE micro-batch input-sharing optimization. Internal/experimental — leave at the default unless investigating an SM12x MoE regression. |
| `FLASHINFER_B12X_FORCE_MOE_W4A16` | unset | `flashinfer/fused_moe/cute_dsl/blackwell_sm12x/moe_dispatch.py` | `1` forces the SM12x MoE dispatcher onto the W4A16 path with BF16 activations. Internal/experimental — used to reproduce W4A16-specific issues. |
| `FLASHINFER_B12X_STATIC_COMPACT_CUTOVER_PAIRS` | shape/device-dependent | `flashinfer/fused_moe/cute_dsl/blackwell_sm12x/moe_dispatch.py` | Override the routed-pair threshold between static and dynamic MoE for device/workload tuning. Set before first use; the decision is cached. A nonnegative integer, with `0` disabling the static band. Legacy aliases, in decreasing precedence: `B12X_STATIC_COMPACT_CUTOVER_PAIRS`, `B12X_DYNAMIC_STATIC_CUTOVER_PAIRS`, `B12X_LEVEL10_STATIC_CUTOVER_PAIRS`. |
| `FLASHINFER_TACTICS_BLOCKLIST` | unset | `flashinfer/autotuner/autotuner.py` | Path to a JSON tactics-blocklist file generated by `flashinfer tactics-blocklist generate` (or `python -m flashinfer tactics-blocklist generate`). When set, the autotuner loads the file at startup and skips any kernel tactics listed as invalid for the current GPU/driver environment, preventing hang or crash on known-bad tactics. |
| `FLASHINFER_CUDNN_PREFILL_SHAPE_OVERRIDE` | `1` | `flashinfer/cudnn/prefill.py` (ragged token-indptr path) | `0` makes the cuDNN ragged prefill graph declare the exact `(batch, max_seq_len)` per call instead of building one override-enabled graph per length class (declared max_len 128 or 65536+, batch 4096+) and passing the real shapes at execute time via cuDNN's shape override. With matching FE 1.31+ Python/native support, BF16 SM100 MLA prefixes (D192/V128, equal 4–64 heads, B1–4, Q64–1024, KV2K–32K at least 4x Q) use bounded power-of-two batch/Q/KV classes and a packed-Q capacity instead; this preserves reuse within each class while exposing useful plan-time occupancy and workspace bounds. Crossing a class builds one cached plan. Override exists because cuDNN finalizes one execution plan per declared shape (55-70 ms, ~1 s with a kernel compile) and serving changes the shape every step; needs cuDNN 9.22+ and cudnn-frontend 1.29+, otherwise the exact declaration is used. Disable only to bisect a suspected override regression. |
| `FLASHINFER_ALLOW_CUDNN_PLAN_BUILD_IN_CAPTURE` | `0` | `flashinfer/gemm/gemm_base.py` (read once at import) | `1` downgrades the refusal to build a cuDNN execution plan during CUDA graph capture from a `CudnnPlanBuildInCaptureError` to a warning. Set it only if you have verified your cuDNN version tolerates plan build under capture; the supported alternative is to run each shape once eagerly outside the capture region, which populates the per-shape plan cache the captured call then reuses. |
| `FLASHINFER_DECODE_AUTO_CUDNN` | unset | `flashinfer/decode.py` (`_auto_decode_prefers_cudnn`) | `BatchDecodeWithPagedKVCacheWrapper(backend="auto")`: unset applies the measured rule (SM100/SM103 with cudnn-frontend 1.30+, fp16/bf16 head_dim 128, GQA group dividing 128, `2 <= q_len_per_req <= 4` with `q_len_per_req * group` in [32, 128] rows per CTA and `batch_size * num_kv_heads >= 64` CTAs, no RoPE / soft-cap / window resolve to `cudnn`, whose default SM100 engine runs those multi-token rows on its FROST decode tile at 0.35-0.95x fa2; single-token decode stays on fa2 except head_dim-256 GQA models with group 4 / 8 / 16 and `batch_size * num_kv_heads` in [96, 256], where the d256 decode tile measures at 0.5-0.95x fa2, from 64 under `use_cuda_graph=True` with a caller-owned table on cudnn-frontend 1.31+, which the wrapper tells the graph is replayed (`is_cuda_graph_replay_expected`) so the tile splits KV; multi-token rows without tensor cores, which have no fa2 kernel, take cudnn whenever its decode path can run them); `0` never resolves to cudnn; `1` applies the rule on an older frontend too (benchmarking; the backend engine is ~8x slower there). Under CUDA graphs `auto` takes cudnn only with a caller-owned `block_tables` (the auto-built table cannot grow once captured) or for multi-token rows without tensor cores, and the resolution is frozen after the first plan. A caller-owned table narrower than 2048 tokens keeps fa2 on cudnn-frontend 1.30 (its placement ranks the backend's prefill-class engine first below that; 1.31+ lifts the limit). `auto` takes cudnn only where the frontend's FROST engines can run (its own CuTe-DSL check: `nvidia-cutlass-dsl` >= 4.7.0, above the 4.6.2 the dependency metadata admits; the `1` override and the no-fa2 route are not gated), and callers passing `fixed_split_size` / `disable_split_kv` (fa2's batch-invariance controls) keep fa2 wherever fa2 can serve the rows. `wrapper.resolved_backend` reports the choice. |
| `FLASHINFER_CAKE_SSD_CHUNK_PARALLEL` | `auto` | `flashinfer/mamba/cake_ssd_combined.py` (`SSDCombined(backend="cake")`, read on every call) | Chooses between the two exact Mamba2 SSD prefill program families of the Cake backend: `auto` applies the calibrated rule (`chunk_parallel_selected`: the chunk-parallel family only when the call has fewer `(sequence, head)` work items than SMs, its per-tile S/H workspace fits in 256 MiB and its predicted main-kernel time beats the exact scan's by the selection margin), `always` forces the chunk-parallel family, `never` forces the exact scan. Both families run the same exact-scan arithmetic. Any other value raises `ValueError`. |
| `FLASHINFER_NVFP4_4OVER6` | unset | `flashinfer/quantization/nvfp4_quantization_utils.py`, `csrc/nv_internal/cpp/common/envUtils.cpp` | **Legacy.** `1` enables the NVFP4 "4over6" block-scale search. Standard NVFP4 gives each 16-element block the single E4M3 scale implied by its absolute max (`amax / 6`, because 6 is the E2M1 max). 4over6 additionally tries the 1.5x tighter `amax / 4`, quantizes the block both ways, dequantizes both, and keeps whichever reconstructs the block with less error — a little more quantizer work for lower quantization error. Requires fp16/bf16 input and E4M3 (not UE8M0) block scales. Consulted **only** when the public `nvfp4_4over6=` parameter is omitted (its default); a `FutureWarning` is emitted only when this variable actually enables 4over6 (unset or `0` stays silent). An explicit `None` (off) or `NVFP44Over6Config(...)` ignores this variable and the three below, with no per-field merge. Prefer the parameter. |
| `FLASHINFER_NVFP4_4OVER6_ERR_MODE` | `MAE` | `flashinfer/quantization/nvfp4_quantization_utils.py`, `csrc/nv_internal/cpp/common/envUtils.cpp` | **Legacy.** Metric used to choose between the two 4over6 candidates: `MAE` (sum of absolute errors) or `MSE` (sum of squared errors). The `amax / 4` candidate buys resolution for the bulk of the block by clipping its largest magnitudes, and MSE penalizes those few large residuals much harder than MAE does, so MSE selects it less often — prefer MSE when a handful of outlier activations dominate downstream accuracy. Read only when `FLASHINFER_NVFP4_4OVER6=1` and `nvfp4_4over6=` is omitted. Parameter spelling: `NVFP44Over6Config(err_mode="MSE")`. |
| `FLASHINFER_NVFP4_4OVER6_ERR_USE_FAST_MATH` | unset | `flashinfer/quantization/nvfp4_quantization_utils.py`, `csrc/nv_internal/cpp/common/envUtils.cpp` | **Legacy.** `1` evaluates the 4over6 candidate error in fp16 (a packed `f16x2` dequant of each candidate) instead of exactly in fp32. Cheaper, and on near-ties it can pick the other candidate, so it is a distinct recipe rather than a pure speed knob. Same gating as `FLASHINFER_NVFP4_4OVER6_ERR_MODE`. Parameter spelling: `NVFP44Over6Config(err_use_fast_math=True)`. |
| `FLASHINFER_NVFP4_4OVER6_E4M3_USE_256` | unset | `flashinfer/quantization/nvfp4_quantization_utils.py`, `csrc/nv_internal/cpp/common/envUtils.cpp` | **Legacy.** `1` treats 256, not the full-range 448, as the top of the E4M3 block-scale range: the per-tensor global scale becomes `256 * 6 / amax` (per-token: `1 / (256 * 6)`) and the kernel's error normalization follows. This is what keeps the 1.5x-larger `amax / 4` candidate representable — `256 * 1.5 = 384` still fits E4M3, whereas a 448-based scale would need `672` and saturate for blocks near the tensor amax. The global scale **must** be built from the same recipe (`make_nvfp4_global_scale(..., nvfp4_4over6_config=...)`); a mismatch silently rescales the whole tensor. Same gating as `FLASHINFER_NVFP4_4OVER6_ERR_MODE`. Parameter spelling: `NVFP44Over6Config(e4m3_max=256)`. |
| `FLASHINFER_DISABLE_FP4_QUANT_FAST_MATH` | unset | `flashinfer/quantization/kernels/nvfp4_quantize.py`, `csrc/nv_internal/cpp/common/envUtils.cpp` | `1` makes the NVFP4 quantizer derive its block/global scales with IEEE round-to-nearest arithmetic (`fdiv_rn`, `nvfp4_scale_from_amax_rn`) instead of the default fast approximate reciprocal (`rcp_approx_ftz`). Slower, but reproduces a reference implementation bit-exactly — use it when bisecting an NVFP4 accuracy mismatch. The flag is part of the CuTe-DSL specialization name (`_nofastmath` suffix), so toggling it compiles a separate kernel rather than reusing the cached one. The C++ side (`getEnvDisableFP4QuantFastMath()`) also accepts TRT-LLM's `TRTLLM_DISABLE_FP4_QUANT_FAST_MATH` as an alias. |

##### Experimental KDA Output-Only Decode Tuning

Low-level overrides for the output-only / RecoverSSM-verify KDA decode
kernels (``flashinfer/kda_kernels/kda_decode_wy_output_only.py``); intended
for benchmarking and bring-up.

| Variable | Default | Effect |
|----------|---------|--------|
| `FLASHINFER_KDA_OO_REC_MAX_BH` | `192` | WY-vs-recurrent auto-dispatch threshold on `B * HV` at `T <= 2` (a quarter of it at `T` in (2, 4]). |
| `FLASHINFER_KDA_OO_REC_KS` | batch policy | Override the grouped recurrent backend's K-slices per state column. |
| `FLASHINFER_KDA_OO_REC_VSPLIT` | batch policy | Override the grouped recurrent backend's value-column split per CTA. |
| `FLASHINFER_KDA_OO_MBP` | grid policy | Override the WY kernel's minimum-blocks-per-SM launch-bound hint. |

##### Experimental Packed-KDA Decode Tuning

These low-level overrides are for benchmarking and kernel bring-up. Production
users should normally leave the batch-size policy at its defaults.

| Variable | Default | Effect |
|----------|---------|--------|
| `FLASHINFER_KDA_T1_FAST_PATH` | `1` | `0` disables routing eligible recurrent KDA T=1 calls to the packed CuTe-DSL decode kernel. |
| `FLASHINFER_PACKED_KDA_TILE_V` | batch policy | Override the V-dimension tile width with a supported packed-KDA tile size. |
| `FLASHINFER_PACKED_KDA_ILP` | batch policy | Override the number of rows processed per instruction-level-parallel group. |
| `FLASHINFER_PACKED_KDA_THREADS` | batch policy | Override CTA threads; the value divided by 16 must select a valid group count. |
| `FLASHINFER_PACKED_KDA_EVICT_FIRST` | batch policy | Set to `1` or `0` to override the policy's first-use cache-eviction hint. |
| `FLASHINFER_PACKED_KDA_CHUNKR` | policy-derived | Override rows per staged state-pool chunk. |
| `FLASHINFER_PACKED_KDA_STAGES` | batch policy | Override cp.async pipeline depth; use `0` for register prefetch or at least `2` for pipelining. |
| `FLASHINFER_PACKED_KDA_NO_CPASYNC` | `0` | `1` forces the register-prefetch kernel instead of cp.async staging. |
| `FLASHINFER_PACKED_KDA_PERSIST` | `0` | `1` enables the persistent-CTA schedule when alignment and pipeline constraints permit it. |
| `FLASHINFER_PACKED_KDA_TMA` | `0` | `1` uses TMA state-pool reads when the staged kernel is active; mutually exclusive with bulk stores. |
| `FLASHINFER_PACKED_KDA_BULK` | `0` | `1` uses bulk stores for an eligible staged kernel; mutually exclusive with TMA reads. |
| `FLASHINFER_PACKED_KDA_PRIVRING` | `1` | `1` selects private-ring synchronization when neither TMA nor bulk-store mode is active. |
| `FLASHINFER_PACKED_KDA_MAXRREG` | `0` | Override the generated kernel's maximum-register tuning hint (`0` leaves it unspecified). |
| `FLASHINFER_PACKED_KDA_L2HINT` | `256` | Override the generated kernel's L2-prefetch-size hint. |
| `FLASHINFER_PACKED_KDA_HPC` | `1` | Override heads combined per CTA for eligible staged whole-head tiles. |
| `FLASHINFER_PACKED_KDA_L2POL` | `0` | Override the experimental L2 cache-policy selector. |
| `FLASHINFER_PACKED_KDA_MINBLOCKS` | `0` | Override the generated kernel's minimum-blocks-per-SM launch-bound hint. |

## Development Workflow

### Typical Development Loop

1. Edit kernel code in `include/flashinfer/some_kernel.cuh`
2. Run test: `pytest tests/test_some_kernel.py::test_specific_case`
3. FlashInfer detects changes and recompiles automatically
4. No `pip install` needed!

### Modifying Existing Kernels

- **Kernel templates**: `include/flashinfer/**/*.cuh` - Changes picked up on next JIT compile
- **Launcher code**: `csrc/*.cu` - May need changes if adding new template parameters
- **Jinja templates**: `csrc/*.jinja` - Update if adding new config parameters
- **Python API**: `flashinfer/*.py` - Update if changing function signatures

### Creating Pre-compiled Packages

When ready to distribute:

```bash
# Build one binary provider for the target GPU.
export FLASHINFER_JIT_CACHE_PROVIDER_ARCH=9.0a
python -m build --no-isolation --wheel flashinfer-jit-cache-provider

# Build the matching shim with an exact dependency on that provider.
export FLASHINFER_JIT_CACHE_PROVIDER_ARCHS="9.0a"
python -m build --no-isolation --wheel flashinfer-jit-cache
```

The provider build runs `flashinfer/aot.py`, which calls the registered
`gen_*_module()` functions and pre-compiles them for exactly one architecture.
The shim contains metadata and exact dependencies, not compiled kernels. Use
the same `FLASHINFER_LOCAL_VERSION` and `FLASHINFER_DEV_RELEASE_SUFFIX` values
for every provider and its shim when producing CUDA-specific or nightly wheels.

## Build System Details

- **Build backend**: Custom PEP 517 backend in `build_backend.py`
- **Data directories**: Build creates symlinks for editable installs:
  - `3rdparty/cutlass` → `flashinfer/data/cutlass`
  - `csrc` → `flashinfer/data/csrc`
  - `include` → `flashinfer/data/include`
- **Version**: Generated in `flashinfer/_build_meta.py` from `version.txt`

## External Integrations

### TVM-FFI: Cross-Language Unified ABI

FlashInfer uses **TVM-FFI** (Apache TVM's Foreign Function Interface) for bindings, which provides a **cross-language unified ABI**. This means:

- **Not limited to PyTorch**: The same compiled kernels can be used from multiple frameworks
- **Language agnostic**: Bindings can be created for Python, C++, Rust, etc.
- **Type-safe marshaling**: Automatic tensor/array conversion between languages
- **Export syntax**: Use `TVM_FFI_DLL_EXPORT_TYPED_FUNC(name, func)` to expose C++ functions

While FlashInfer currently provides PyTorch bindings, the underlying kernels are framework-agnostic thanks to TVM-FFI.

### Other Integrations

- **PyTorch Custom Ops**: `torch.library` for `torch.compile()` and CUDA graph support
- **Ninja Build**: Direct ninja generation, no CMake complexity

## Supported GPU Architectures

FlashInfer supports NVIDIA SM75, SM80, SM86, SM89, SM90, SM100, SM103, SM110, SM120, and SM121.

## Release Versioning

FlashInfer follows a "right-shifted" versioning scheme (`major.minor.patch[.post1]`):

- **major**: Architectural milestone and/or incompatible API changes (similar to PyTorch 2.0)
- **minor**: Significant backwards-compatible new features
- **patch**: Small backwards-compatible features (new kernels, new SM support) and backwards-compatible bug fixes
- **post1**: Optional suffix for quick follow-up release with just backwards-compatible bug fixes

## External Documentation Resources

When working with FlashInfer's dependencies and tools, refer to these official documentation sources:

### Core Dependencies

- **TVM-FFI**: Apache TVM's Foreign Function Interface
  - Documentation: <https://tvm.apache.org/ffi/>
  - Package: `apache-tvm-ffi` (<https://pypi.org/project/apache-tvm-ffi/>)
  - Use for: Understanding FFI export syntax, cross-language bindings

- **CUTLASS**: NVIDIA's CUDA Templates for Linear Algebra Subroutines
  - **Recommended**: Read source code directly in `3rdparty/cutlass/` (documentation is often outdated)
  - Repository: <https://github.com/NVIDIA/cutlass>
  - Use for: GEMM kernel implementations, tensor core operations

- **CuTe (CUTE DSL)**: CUTLASS's Cute Layout and Tensor DSL
  - Documentation: <https://docs.nvidia.com/cutlass/media/docs/pythonDSL/cute_dsl.html>
  - **Tip**: Add `.md` to get Markdown format: <https://docs.nvidia.com/cutlass/media/docs/pythonDSL/cute_dsl.html.md>
  - The Cute DSL kernels rely on Python modules from the `nvidia-cutlass-dsl` pip package, not to be confused with Python modules in the `3rdparty/cutlass` submodule
  - Tutorial: <https://github.com/NVIDIA/cutlass/tree/main/examples/python/CuTeDSL>

- **PTX ISA (Parallel Thread Execution)**: NVIDIA's PTX instruction set documentation
  - Documentation: <https://docs.nvidia.com/cuda/parallel-thread-execution/>
  - **Index/Table of Contents**: <https://docs.nvidia.com/cuda/parallel-thread-execution/index.html.md>
  - **Tip**: Add `.md` to any page URL to get Markdown format
  - Use for: Low-level instruction details, new GPU architecture features, inline PTX assembly

### When to Consult These Docs

- **Understanding new GPU architecture features** → Check PTX ISA documentation for latest instruction details
- **Working on FFI bindings** → Check TVM-FFI docs for export patterns and type marshaling
- **Implementing Tensor-Core kernels using CUTLASS** → Read source code in `3rdparty/cutlass/`
- **Using tensor layouts or warp-level operations** → Refer to CuTe documentation
- **Writing inline PTX assembly** → Consult PTX ISA for instruction syntax and semantics

These dependencies are included in FlashInfer's `3rdparty/` directory or `requirements.txt`.

### Some final suggestions for all AI agents

> Because practical engineering involves the accumulated experience of trial and error, match the coding style, efficiency, complexity, verbosity, and defensiveness by learning from existing code as much as possible—this document contains many pointers on where to find examples. Document intentional departures with rationale. Mentioning "AI-assisted" in the git commit message is good transparency. For performance-critical hot paths, leave justification for the special algorithmic choices and other potential alternatives in a comment for review.

**Keep documentation in sync with code changes:** When modifying code that is referenced in this document or in `.claude/skills/`, update the corresponding documentation immediately. This includes:
- Important infrastructure changes (e.g., `@flashinfer_api`, `@backend_requirement`, TVM-FFI macros) → Update examples in `CLAUDE.md` and relevant skill files
- New patterns or conventions → Document them for future reference
- Deprecated approaches → Remove or mark as deprecated in docs
- New error handling patterns, macros, or utilities → Add to relevant skill tutorials