For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo module

dispatch_fused_bias_residual

Dispatcher for fused matmul + bias/residual (mo.composite.matmul_add).

Kept separate from matmul_dispatch_sm100 so the generic matmul dispatch does not have to thread an epilogue tensor through every kernel. The bias/residual is honored one of two mutually exclusive ways:

• Native: the SM100 blackwell GEMM applies it via its TMA epilogue load (the tensor is passed, no lambda). Used for GEMM-shaped problems (M>1, N>1, bf16, 16B-aligned N/K) -- the fast prefill path.
• Fallback: the bias/residual is wrapped as a normal elementwise (store) epilogue and the generic matmul_dispatch_sm100 is reused. GEMV (M=1/N=1), small-MN, and vendor (cuBLAS) all already apply an elementwise epilogue, so correctness is universal. No tensor is passed, so there is no double-add.

comptime values

logger

comptime logger = Logger(stdout, prefix=String(""), source_location=False)

Functions

• ​fused_bias_residual_matmul_dispatch_sm100:

Was this page helpful?

Thank you! We'll create more content like this.

Thank you for helping us improve!

😔 What went wrong?

Some code doesn’t work

It includes inaccurate information

It's missing information I need

It was difficult to understand

Other

Submit