Commits

Breadcrumbs

History for

onmain

User selector

All users

Datepicker

All time

Commit history

Commits on Aug 18, 2026

• Fix ROCm 7.14 atomicAdd clash and the missing DTensor _utils stub

  Show description for 2fc5f43
  

  ThePie88

  and

  claude

  committed
  2fc5f43View commit details

  Copy full SHA for 2fc5f43

  View code at this pointBrowse repository at this point

Commits on Aug 9, 2026

• Remove hardcoded machine paths from the build scripts so a fresh clone builds unedited

  Show description for 7c16fdd
  

  ThePie88

  and

  claude

  committed
  7c16fddView commit details

  Copy full SHA for 7c16fdd

  View code at this pointBrowse repository at this point

Commits on Jul 3, 2026

• CK ck_tile FMHA (WMMA) prefill for gfx1100, wired into vLLM (1.09-1.39x TTFT, grows with context)

  Show description for d5396f0
  

  ThePie88

  and

  claude

  committed
  d5396f0View commit details

  Copy full SHA for d5396f0

  View code at this pointBrowse repository at this point

Commits on Jul 2, 2026

• MoE-decode gate_up GEMV: @triton.autotune (2x isolated; +1.3% gemma / ~0% ERNIE e2e)

  Show description for fc6620f
  

  ThePie88

  and

  claude

  committed
  fc6620fView commit details

  Copy full SHA for fc6620f

  View code at this pointBrowse repository at this point

• Flash-layout decode kernel (correct, but native attention only beats Triton at head 256); ERNIE in README benchmark

  Show description for 814421b
  

  ThePie88

  and

  claude

  committed
  814421bView commit details

  Copy full SHA for 814421b

  View code at this pointBrowse repository at this point

Commits on Jul 1, 2026

• Native paged-attention decode kernel + sliding mask; ablation finds attention is the decode lever; ERNIE clean-VRAM bench; AITER-on-RDNA3 intent

  Show description for 5d79222
  

  ThePie88

  and

  claude

  committed
  5d79222View commit details

  Copy full SHA for 5d79222

  View code at this pointBrowse repository at this point

• Native _rocm_C skinny GEMM (wvSplitK) built for Win ROCm -> 50 -> 57 tok/s decode

  Show description for 5e463fa
  

  ThePie88

  and

  claude

  committed
  5e463faView commit details

  Copy full SHA for 5e463fa

  View code at this pointBrowse repository at this point

• bf16 skinny M=1 GEMV for the dense MLP (opt-in) + finding: decode is latency-bound

  Show description for baaa0be
  

  ThePie88

  and

  claude

  committed
  baaa0beView commit details

  Copy full SHA for baaa0be

  View code at this pointBrowse repository at this point

• moe_decode: gather-in-kernel W4 GEMV -> 40 -> 50 tok/s decode on gemma-4-26B

  Show description for fd6cf79
  

  ThePie88

  and

  claude

  committed
  fd6cf79View commit details

  Copy full SHA for fd6cf79

  View code at this pointBrowse repository at this point

• moe_decode: opt-in loader for the native HIP fused MoE-decode kernel

  Show description for 7c0b10e
  

  ThePie88

  and

  claude

  committed
  7c0b10eView commit details

  Copy full SHA for 7c0b10e

  View code at this pointBrowse repository at this point

• KVarN KV-cache quant on vLLM Win ROCm: correct e2e on gemma-4-26B (+cudagraph 40 tok/s)

  Show description for c97e8b6
  

  ThePie88

  and

  claude

  committed
  c97e8b6View commit details

  Copy full SHA for c97e8b6

  View code at this pointBrowse repository at this point

Commits on Jun 30, 2026

• gemma-4-26B MoE (compressed-tensors W4A16) on vLLM Win ROCm: working + 1.1->50.4 tok/s

  Show description for b031c6c
  

  ThePie88

  and

  claude

  committed
  b031c6cView commit details

  Copy full SHA for b031c6c

  View code at this pointBrowse repository at this point

• Hand-written HIP W4 GEMV (buffer_load dwordx4 + unroll + split-K), opt-in

  Show description for 718adec
  

  ThePie88

  and

  claude

  committed
  718adecView commit details

  Copy full SHA for 718adec

  View code at this pointBrowse repository at this point

• Autotune the AWQ M=1 GEMV: 14B-AWQ 37.7 -> 50.9 tok/s (4.2x over conch)

  Show description for b0fb9db
  

  ThePie88

  and

  claude

  committed
  b0fb9dbView commit details

  Copy full SHA for b0fb9db

  View code at this pointBrowse repository at this point

• Custom AWQ-uint4 M=1 GEMV + GPTQ/AWQ benchmarks (14B-AWQ 12.2 -> 37.7 tok/s)

  Show description for edd7166
  

  ThePie88

  and

  claude

  committed
  edd7166View commit details

  Copy full SHA for edd7166

  View code at this pointBrowse repository at this point

• Native HIP kernels + inductor: decode 11.4 -> 39.9 tok/s on gfx1100

  Show description for 11969a9
  

  ThePie88

  and

  claude

  committed
  11969a9View commit details

  Copy full SHA for 11969a9

  View code at this pointBrowse repository at this point

• Initial commit: vLLM on native Windows + AMD ROCm (RDNA3)

  Show description for 3a3b2c9
  

  ThePie88

  and

  claude

  committed
  3a3b2c9View commit details

  Copy full SHA for 3a3b2c9

  View code at this pointBrowse repository at this point

Loading