Optimize search space and upgrade Image to 0.19.0 for MiniMax-M2.5 (#1003)

* Add TP2EP2 for minimaxm2.5-fp8-mi355x-vllm

Fewer GPUs means less inter-GPU communication overhead, and MoE
expert parallelism across 2 GPUs is very efficient for this model.

* Optimize config for minimaxm2.5-fp8-mi355x-vllm

* Update perf-changelog for minimaxm2.5-fp8-mi355x-vllm

* Upgrade minimaxm2.5-fp8-mi355x-vllm Image to v0.19.0

Enable FP8 KV cache + AITER FA for minimaxm2.5-fp8-mi355x-vllm

* optimize all reduce

* fix pr

* Update perf-chagelog

* Fix the perf-changelog

---------

Co-authored-by: zhutaoyu <zhutaoyu97@gmail.com>

Author: chunfangamd

.github/configs/amd-master.yaml

@@ -334,7 +334,7 @@ kimik2.5-fp4-mi355x-atom:
     - { tp: 4, conc-start: 4, conc-end: 128 }
 
 minimaxm2.5-fp8-mi355x-vllm:
-  image: vllm/vllm-openai-rocm:v0.18.0
+  image: vllm/vllm-openai-rocm:v0.19.0
   model: MiniMaxAI/MiniMax-M2.5
   model-prefix: minimaxm2.5
   runner: mi355x
@@ -345,15 +345,15 @@ minimaxm2.5-fp8-mi355x-vllm:
   - isl: 1024
     osl: 1024
     search-space:
-    - { tp: 2, conc-start: 4, conc-end: 64 }
-    - { tp: 4, conc-start: 4, conc-end: 64 }
-    - { tp: 8, ep: 8, conc-start: 32, conc-end: 256 }
+    - { tp: 2, ep: 2, conc-start: 2, conc-end: 512 }
+    - { tp: 4, ep: 4, conc-start: 4, conc-end: 256 }
+    - { tp: 8, ep: 8, conc-start: 2, conc-end: 2 }
   - isl: 8192
     osl: 1024
     search-space:
-    - { tp: 2, conc-start: 4, conc-end: 64 }
-    - { tp: 4, conc-start: 4, conc-end: 64 }
-    - { tp: 8, ep: 8, conc-start: 32, conc-end: 256 }
+    - { tp: 2, ep: 2, conc-start: 2, conc-end: 256 }
+    - { tp: 4, ep: 4, conc-start: 4, conc-end: 512 }
+    - { tp: 8, ep: 8, conc-start: 2, conc-end: 2 }
 
 minimaxm2.5-fp8-mi355x-atom:
   image: rocm/atom:rocm7.2.1-ubuntu24.04-pytorch2.9.1-atom0.1.2

benchmarks/single_node/minimaxm2.5_fp8_mi355x.sh

@@ -25,6 +25,7 @@ if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
 fi
 
 export VLLM_ROCM_USE_AITER=1
+export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
 
 SERVER_LOG=/workspace/server.log
 PORT=${PORT:-8888}
@@ -49,8 +50,10 @@ vllm serve $MODEL --port $PORT \
 $EP \
 --gpu-memory-utilization 0.95 \
 --max-model-len $MAX_MODEL_LEN \
+--kv-cache-dtype fp8 \
 --block-size=32 \
 --no-enable-prefix-caching \
+--attention-backend "ROCM_AITER_FA" \
 --trust-remote-code > $SERVER_LOG 2>&1 &
 
 SERVER_PID=$!

perf-changelog.yaml

@@ -1244,4 +1244,18 @@
     - "Remove ISL 1024 / OSL 8192 seq-len config"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/947
 
+- config-keys:
+    - minimaxm2.5-fp8-mi355x-vllm
+  description:
+    - "Optimize MiniMax-M2.5 FP8 MI355X vLLM search-space"
+    - "Add tp2 ep2 search-space entries (conc 2-256) for all seq-len configs"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1002
 
+- config-keys:
+    - minimaxm2.5-fp8-mi355x-vllm
+  description:
+    - "Optimize MiniMax-M2.5 FP8 MI355X vLLM search-space"
+    - "Add tp2 ep2 search-space entries (conc 2-256) for all seq-len configs"
+    - "Upgrade vLLM image to v0.19.0"
+    - "Enable FP8 KV cache + AITER FA for minimaxm2.5-fp8-mi355x-vllm"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1003