See the compiler think.
Every pass, in completion order, with the full IR before and after. Browse pinned examples or bring your own trace manifest.
Compile and inspect
Use the same pass viewer for a kernel you just wrote.
Trace catalog
versioned pass-trace/v1 · snapshots load on demand
add_kernel
The canonical masked vector add: a compact baseline for pointer, mask, and layout lowering.
softmax
Fused softmax on AMD: row reductions, numerical stabilization, and layout conversion.
softmax
The same fused softmax shape on CUDA, exposing target-specific lowering and LLVM passes.
dropout_kernel
Seeded low-memory dropout with counter-based random values, masking, scaling, and stores.
layer_norm_kernel
Layer normalization with mean/variance reductions and an affine epilogue.
matmul_kernel
Grouped, tiled matrix multiplication with dot lowering and AMD matrix-instruction paths.
fused_attention
Fused attention lowering with scheduling, coalescing, and LLVM conversion.
memcpy_layout_kernel
A layout-declared memcpy that makes Gluon auto-encoding resolution visible.
async_copy_kernel
Ampere Gluon async-copy pipeline: shared memory allocation, commit/wait, and lowering.
_rms_norm_forward_kernel
LinkedIn Liger production RMSNorm forward: Llama/Gemma casting modes, cached rstd, and affine epilogue.
_swiglu_forward_kernel
Liger fused SwiGLU MLP activation: SiLU(gate) * up with float32 sigmoid and gate multiplier.
_softmax_single_block_forward_kernel
Liger single-block row softmax: the production fast path when a row fits in one block.
v9_beyond_hotloop
AMD ROCm peak FP16 GEMM: XCD-aware PID remapping, async-copy MFMA pipeline, and AGPR allocation.
_swiglustep_and_mul_kernel
vLLM GPT-OSS-style activation: clamped SiLU(gate) × up from a split gate/up row.
softcap_out_kernel
SGLang serving soft-cap: portable tanh-style softcap used on logits and attention scores.
_gemma_rmsnorm_kernel
AMD ATOM fused residual + Gemma RMSNorm with weight offset (1 + w) and persistent rows.
_softmax_kernel_online
AMD AITER two-pass online softmax: streaming max/sum over wide rows for Instinct GPUs.
rms_norm_kernel
NVIDIA TensorRT-LLM auto-deploy RMSNorm with constexpr stride/width and float32 weight multiply.
merge_attn_states_kernel
vLLM split-KV attention merge: combines prefix/suffix partial outputs and LSE values.
merge_attn_states_kernel
vLLM split-KV attention merge: combines prefix/suffix partial outputs and LSE values.
_flash_decode_stage2_kernel
TensorRT-LLM flash-decode stage-2: reduces split-K partial attention outputs by LSE weights.
_flash_decode_stage2_kernel
TensorRT-LLM flash-decode stage-2: reduces split-K partial attention outputs by LSE weights.
_fused_topk_softmax_kernel
TensorRT-LLM fused MoE top-k + softmax: selects experts from raw logits then softmaxes only k.
_fused_topk_softmax_kernel
TensorRT-LLM fused MoE top-k + softmax: selects experts from raw logits then softmaxes only k.
_moe_align_block_size_stage1_kernel
AMD AITER MoE align stage-1: counts tokens per expert before block padding for fused MoE GEMMs.
_moe_align_block_size_stage1_kernel
AMD AITER MoE align stage-1: counts tokens per expert before block padding for fused MoE GEMMs.
moe_fused_mul_sum_kernel
vLLM MoE fused mul+sum: weights expert outputs by router scores and reduces across top-k.
moe_fused_mul_sum_kernel
vLLM MoE fused mul+sum: weights expert outputs by router scores and reduces across top-k.
fused_moe_kernel
vLLM fused MoE expert GEMM: grouped token×expert matmul over sorted/padded token ids.
fused_moe_kernel
vLLM fused MoE expert GEMM: grouped token×expert matmul over sorted/padded token ids.
_router_triton_kernel
SGLang MoE fused gate: sigmoid scoring, biased top-k selection, and optional shared experts.
_router_triton_kernel
SGLang MoE fused gate: sigmoid scoring, biased top-k selection, and optional shared experts.
_moe_router_histogram_kernel
Liger SonicMoE-style router histogram: counts (token, k) assignments per expert per tile.
_moe_router_histogram_kernel
Liger SonicMoE-style router histogram: counts (token, k) assignments per expert per tile.