Code

[Models] Implement support for gemma#219

Code

Merged

sunggg merged 10 commits intobatch-servingoctoml/mlc-llm:batch-servingfrom lunderberg/dev_gemmaoctoml/mlc-llm:lunderberg/dev_gemmaCopy head branch name to clipboardFeb 22, 2024

+147-31

Commits

Commits on Feb 22, 2024

• model compiles

  

  masahi

  authored and

  Lunderberg

  committed
  6f20f25

  Copy full SHA for 6f20f25

• use embedding weight for lm_head

  

  masahi

  authored and

  Lunderberg

  committed
  0379f5a

  Copy full SHA for 0379f5a

• support gemma style rms

  

  masahi

  authored and

  Lunderberg

  committed
  eb3058f

  Copy full SHA for eb3058f

• fix head_dim

  

  masahi

  authored and

  Lunderberg

  committed
  107e408

  Copy full SHA for 107e408

• runs but output garbage

  

  masahi

  authored and

  Lunderberg

  committed
  f3a9f6a

  Copy full SHA for f3a9f6a

• use gelu activation

  

  masahi

  authored and

  Lunderberg

  committed
  41af657

  Copy full SHA for 41af657

• fix

  

  masahi

  authored and

  Lunderberg

  committed
  8c68a3d

  Copy full SHA for 8c68a3d

• Fix, use hasattr to check if head_dim is available

  Lunderberg

  committed
  5ccdd57

  Copy full SHA for 5ccdd57

• Bugfix, use the weight_offset at all locations for LlamaRMSNorm

  Lunderberg

  committed
  c516429

  Copy full SHA for c516429

• Bug-workaround, cutlass.rms_norm requires weight_offset of zero

  Lunderberg

  committed
  353edbf

  Copy full SHA for 353edbf