Repository navigation

[AMD][MI355X] Kimi-K3 vLLM AgentX FP8 per-channel linear quantization-config / [AMD][MI355X] Kimi-K3 vLLM AgentX 新增 FP8 per-channel linear quantization-config - #3627

Closed

seungrokj wants to merge 12 commits into

mainfrom

amd/agentx_k3_vllm_1001

Closed

[AMD][MI355X] Kimi-K3 vLLM AgentX FP8 per-channel linear quantization-config / [AMD][MI355X] Kimi-K3 vLLM AgentX 新增 FP8 per-channel linear quantization-config#3627
seungrokj wants to merge 12 commits into
mainfrom
amd/agentx_k3_vllm_1001

Commits on Oct 1, 2026
- [AMD][MI355X] Kimi-K3 vLLM AgentX: FP8 per-channel linear quantization-config
  
  seungrokj
  and
  claude
  committed
  db21861
- Set perf-changelog pr-link to #3627
  
  seungrokj
  and
  claude
  committed
  1a911e1
- [AMD][MI355X] Kimi-K3 vLLM AgentX: concurrency 10-16 and add 75
  
  seungrokj
  and
  claude
  committed
  f7bcd50
- Add Kimi-K3 vLLM c16 and c75 recipe variants for the new sweep
  
  seungrokj
  and
  claude
  committed
  b2cc028
- [AMD][MI355X] Kimi-K3 vLLM AgentX: load with safetensors instead of fastsafetensors
  
  seungrokj
  and
  claude
  committed
  4425b4e
- Merge branch 'main' into amd/agentx_k3_vllm_1001
  seungrokj
  authored
  a89e81f
- perf(amd): keep non-beneficial Kimi-K3 linears in BF16 under online FP8
  
  seungrokj
  and
  claude
  committed
  7700e47
- chore(amd): narrow Kimi-K3 MI355X vLLM AgentX sweep to c4 and dcp8 c80
  
  seungrokj
  and
  claude
  committed
  519c72a
- fix(amd): capture Kimi-K3 c8 CUDA graphs up to 2x CONC x (1 + drafts)
  
  seungrokj
  and
  claude
  committed
  026ac29
- chore(amd): add Kimi-K3 MI355X vLLM c80 dcp8 variant and sweep only c80
  
  seungrokj
  and
  claude
  committed
  22c0cb8
Commits on Oct 2, 2026
- chore(amd): restore full Kimi-K3 MI355X vLLM AgentX sweep
  
  seungrokj
  and
  claude
  committed
  396a6c4
- Merge branch 'main' into amd/agentx_k3_vllm_1001
  seungrokj
  authored
  7ee192f