Skip to content

roofline: emit per-launch topology edges for graph reconstruction#66

Open
roberteg16 wants to merge 2 commits into
gfx11from
rogarcia.roofline-graph-topology
Open

roofline: emit per-launch topology edges for graph reconstruction#66
roberteg16 wants to merge 2 commits into
gfx11from
rogarcia.roofline-graph-topology

Conversation

@roberteg16

Copy link
Copy Markdown

Record per-invocation storage-id identity (out/in) so the consumer can link each launch to its producer via last-writer-wins on storage roots, reconstructing the execution graph. Fused spans elide internal tensors.

Counter part of: https://gitenterprise.xilinx.com/FaaSApps/rocm-scripts/pull/781

Record per-invocation storage-id identity (out/in) so the consumer can
link each launch to its producer via last-writer-wins on storage roots,
reconstructing the execution graph. Fused spans elide internal tensors.

Co-Authored-By: Claude Opus 4 (1M context) <[email protected]>
…_ids

Each launch's topology now carries, 1:1 with in_storage_ids, the source operand's
tensor name (in_names), dtype (in_types) and shape (in_ne). The name makes
weight-vs-input unambiguous for the consumer -- a model weight is named e.g.
"blk.0.ffn_gate.weight" / "*.bias", a dynamic input "inp_pos" / "cache_k_l0" -- and
the per-operand type/shape are correct even for a fused span's non-head operands,
which the shared (head-only) geometry record does not describe. The name is taken
from the storage root so a view/reshape of a weight keeps the weight's name.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant