Skip to content

Enable weightfree export with dynamo for Kimi K2.5 - #1237

Draft
quic-mamta wants to merge 48 commits into
quic:mainfrom
quic-mamta:kimi_dynamo
Draft

Enable weightfree export with dynamo for Kimi K2.5#1237
quic-mamta wants to merge 48 commits into
quic:mainfrom
quic-mamta:kimi_dynamo

Conversation

@quic-mamta

@quic-mamta quic-mamta commented Aug 4, 2026

Copy link
Copy Markdown
Contributor

Summary

This PR enables weight-free Dynamo export for Kimi K2.5. It adds a new checkpoint-backed ONNX export path that produces a lightweight ONNX plus weight_spec.json, allowing model weights to remain external instead of being embedded as ONNX initializers.

Changes

  • Add QEfficient.exporter.weight_free package for:
    • weight-free ONNX export
    • weight_spec.json generation/loading
    • checkpoint-backed ORT input loading
    • checkpoint transform pipeline
  • Add use_weight_free_export support in the QEff export/compile flow.
  • Add checkpoint transforms for dtype conversion, MoE expert stacking, fused expert splitting, and GPT-OSS MXFP4 expert dequant/split handling.
  • Register Dynamo custom op support for qefficient::cast_to_uint4.
  • Update Kimi K2.5 Dynamo export paths and examples.
  • Move Kimi loading utilities into the runtime package.
  • Add Kimi K2.5 Dynamo and weight-free validation scripts/tests.
  • Update multimodal/image-text test coverage for Kimi K2.5 flows.

Testing

Added/updated coverage for:

  • Kimi K2.5 Dynamo export/compile/execute flow
  • Kimi K2.5 weight-free Dynamo export flow

mamtsing and others added 30 commits June 19, 2026 16:13
…lQPC and upgrade TF

Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <168400541+quic-mamta@users.noreply.github.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <168400541+quic-mamta@users.noreply.github.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <168400541+quic-mamta@users.noreply.github.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <168400541+quic-mamta@users.noreply.github.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
mamtsing and others added 8 commits July 24, 2026 21:19
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <168400541+quic-mamta@users.noreply.github.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta quic-mamta self-assigned this Aug 4, 2026
@ochougul
ochougul marked this pull request as draft August 4, 2026 16:39
Signed-off-by: Mamta Singh <168400541+quic-mamta@users.noreply.github.com>
Signed-off-by: Mamta Singh <168400541+quic-mamta@users.noreply.github.com>
@quic-mamta
quic-mamta force-pushed the kimi_dynamo branch 4 times, most recently from cc1489b to e1fa3b2 Compare August 5, 2026 11:09
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@vbaddi

vbaddi commented Aug 5, 2026

Copy link
Copy Markdown
Contributor

this is demo specific test pr and will be merged post #1220 is merged.
cc: @quic-rishinr @ochougul

Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <168400541+quic-mamta@users.noreply.github.com>
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.

3 participants