Commits

Branch selector

 main

User selector

All users

All time

Commit History

Commits on Jun 6, 2026

• Run Qwen3-Next quantized in place

  Show description for f9d5694
  

  Duc-Tam Nguyen

  committed
  f9d5694

  Copy full SHA for f9d5694

• Run the Gemma 4 text decoder quantized in place

  Show description for 806ab35
  

  Duc-Tam Nguyen

  committed
  806ab35

  Copy full SHA for 806ab35

• Run the Phi-4 family quantized in place

  Show description for ffefc36
  

  Duc-Tam Nguyen

  committed
  ffefc36

  Copy full SHA for ffefc36

• Run GLM-4 quantized in place

  Show description for d883cbc
  

  Duc-Tam Nguyen

  committed
  d883cbc

  Copy full SHA for d883cbc

• Run quantized Ministral 3 checkpoints in place via the shared primitives

  Show description for bc9a618
  

  Duc-Tam Nguyen

  committed
  bc9a618

  Copy full SHA for bc9a618

• Run quantized Llama checkpoints in place via the shared primitives

  Show description for 403953c
  

  Duc-Tam Nguyen

  committed
  403953c

  Copy full SHA for 403953c

• Run quantized Qwen3 checkpoints in place via the shared primitives

  Show description for 1e1e581
  

  Duc-Tam Nguyen

  committed
  1e1e581

  Copy full SHA for 1e1e581

• Add shared affine quantization primitives for the forward path

  Show description for ead0eff
  

  Duc-Tam Nguyen

  committed
  ead0eff

  Copy full SHA for ead0eff

• compute/models: absorb DeepSeek kv_b_proj into embed_q and unembed_out

  Show description for a00ad8d
  

  Duc-Tam Nguyen

  committed
  a00ad8d

  Copy full SHA for a00ad8d

• compute/models: DeepSeek int4 checkpoint remap, with View and Astype

  Show description for af559e1
  

  Duc-Tam Nguyen

  committed
  af559e1

  Copy full SHA for af559e1

• compute/models: DeepSeek expert stacking and Sanitize key drops

  Show description for 627d65d
  

  Duc-Tam Nguyen

  committed
  627d65d

  Copy full SHA for 627d65d

• mlxgo: add Stack binding

  Show description for 5ae266e
  

  Duc-Tam Nguyen

  committed
  5ae266e

  Copy full SHA for 5ae266e

• mlxgo: add Quantize and Dequantize bindings

  Show description for b057472
  

  Duc-Tam Nguyen

  committed
  b057472

  Copy full SHA for b057472

• compute/models: quantized SwitchGLU expert path via gather_qmm

  Show description for 0024fe2
  

  Duc-Tam Nguyen

  committed
  0024fe2

  Copy full SHA for 0024fe2

• mlxgo: add GatherQMM binding for quantized expert matmul

  Show description for 1193720
  

  Duc-Tam Nguyen

  committed
  1193720

  Copy full SHA for 1193720

• Admit a ragged prompt cohort in one batched prefill

  Show description for 6a3fce9
  

  Duc-Tam Nguyen

  committed
  6a3fce9

  Copy full SHA for 6a3fce9

• Add model-side BatchPrefill for ragged cohort admission

  Show description for cb80568
  

  Duc-Tam Nguyen

  committed
  cb80568

  Copy full SHA for cb80568

• Read cache in qwen3next forward for ragged cohorts

  Show description for b339369
  

  Duc-Tam Nguyen

  committed
  b339369

  Copy full SHA for b339369

• Read cache in deepseek forward for ragged cohorts

  Show description for f66b27d
  

  Duc-Tam Nguyen

  committed
  f66b27d

  Copy full SHA for f66b27d

• Fix off-by-one in the ragged decode attention mask

  Show description for 193e0b5
  

  Duc-Tam Nguyen

  committed
  193e0b5

  Copy full SHA for 193e0b5

• Read cache in gemma forward for ragged cohorts

  Show description for 19800aa
  

  Duc-Tam Nguyen

  committed
  19800aa

  Copy full SHA for 19800aa

• Read the attention mask and rope offset from the cache in the dense forwards

  Show description for 6d07656
  

  Duc-Tam Nguyen

  committed
  6d07656

  Copy full SHA for 6d07656

Commits on Jun 4, 2026

• Carry per-row left padding on the batched cache

  Show description for e9aa889
  

  Duc-Tam Nguyen

  committed
  e9aa889

  Copy full SHA for e9aa889

• Add the left-padded causal masks for ragged batching

  Show description for 070c0d8
  

  Duc-Tam Nguyen

  committed
  070c0d8

  Copy full SHA for 070c0d8

• Decode a synchronized batch in one forward

  Show description for a640d05
  

  Duc-Tam Nguyen

  committed
  a640d05

  Copy full SHA for a640d05

• Extend batched decode to the MoE and hybrid models

  Show description for 9913e00
  

  Duc-Tam Nguyen

  committed
  9913e00

  Copy full SHA for 9913e00

• Add the batched decode forward for the dense models

  Show description for f6c6593
  

  Duc-Tam Nguyen

  committed
  f6c6593

  Copy full SHA for f6c6593

• Drop em-dashes from doc comments

  Show description for 3cabf7a
  

  Duc-Tam Nguyen

  committed
  3cabf7a

  Copy full SHA for 3cabf7a

• Add the Qwen3-Next hybrid forward pass

  Show description for e9af2af
  

  Duc-Tam Nguyen

  committed
  e9af2af

  Copy full SHA for e9af2af

• mlxgo: add exp, logaddexp and repeat for the gated-delta gate

  

  Duc-Tam Nguyen

  committed
  1a4bf07

  Copy full SHA for 1a4bf07

• Add the DeepSeek-V3 forward pass

  Show description for 6c314f2
  

  Duc-Tam Nguyen

  committed
  6c314f2

  Copy full SHA for 6c314f2

• compute/models: SwitchGLU expert MLP forward

  

  Duc-Tam Nguyen

  committed
  f0c1850

  Copy full SHA for f0c1850

• mlxgo: add axis sum and put-along-axis for expert routing

  

  Duc-Tam Nguyen

  committed
  d9cac99

  Copy full SHA for d9cac99

• mlxgo: add gather-matmul, argsort and floor-divide for the expert path

  

  Duc-Tam Nguyen

  committed
  078e753

  Copy full SHA for 078e753

• compute/models: Gemma4 text forward

  Show description for 58508c1
  

  Duc-Tam Nguyen

  committed
  58508c1

  Copy full SHA for 58508c1