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 struct
DenseWeightLoader
struct DenseWeightLoader[b_dtype: DType, accum_dtype: DType = DType.float32]
Direct-DRAM B policy: dense-bf16 and FP8-W8A16.
The B strip is a plain (BK, SG_N) DRAM sub-tile (coalesced via the
col-major (K, N) view for transpose_b=True) fed straight to the native
MMA -- no staging. b_dtype == in_type is the dense case; b_dtype=fp8 is
the W8A16 weight-only case (the only divergence from dense is the operand
dtype). Stateless: pure comptime policy.
Implemented traitsβ
AnyType,
Copyable,
ImplicitlyCopyable,
ImplicitlyDeletable,
Movable,
WeightLoader
comptime membersβ
accum_typeβ
comptime accum_type = accum_dtype
b_typeβ
comptime b_type = b_dtype
needs_smemβ
comptime needs_smem = False
smem_elemsβ
comptime smem_elems = 0
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!