[Fix] Remove MoRI-IO patches from vLLM Disagg benchmarks

.github/configs/amd-master.yaml

@@ -1524,7 +1524,9 @@ dsr1-fp8-mi355x-sglang-disagg-mtp:
           - "DECODE_MTP_SIZE=2"
 
 kimik2.5-fp4-mi355x-vllm-disagg:
-  image: vllm/vllm-openai-rocm:nightly-bf610c2f56764e1b30bc6065f4ceace3d6e59036
+  # TODO(simondanielsson): change to pinned version once https://github.com/vllm-project/vllm/pull/40344
+  # is part of official release, likely 0.22.0.
+  image: vllm/vllm-openai-rocm:nightly-3f0a91bb96f8d72e0498b95c166e817deae14d62
   model: amd/Kimi-K2.5-MXFP4
   model-prefix: kimik2.5
   runner: mi355x-disagg
@@ -1547,7 +1549,6 @@ kimik2.5-fp4-mi355x-vllm-disagg:
           dp-attn: false
           additional-settings:
           - "PREFILL_NODES=1"
-          - "VLLM_MORIIO_CONNECTOR_READ_MODE=1"
         decode:
           num-worker: 2
           tp: 8
@@ -1568,7 +1569,6 @@ kimik2.5-fp4-mi355x-vllm-disagg:
           dp-attn: false
           additional-settings:
           - "PREFILL_NODES=1"
-          - "VLLM_MORIIO_CONNECTOR_READ_MODE=1"
         decode:
           num-worker: 2
           tp: 8
@@ -1578,7 +1578,9 @@ kimik2.5-fp4-mi355x-vllm-disagg:
           - "DECODE_NODES=2"
 
 minimaxm2.5-fp8-mi355x-vllm-disagg:
-  image: vllm/vllm-openai-rocm:nightly-a6682d1d259cca69a9ae737ea5608fbbe7520031
+  # TODO(simondanielsson): change to pinned version once https://github.com/vllm-project/vllm/pull/40344
+  # is part of official release, likely 0.22.0.
+  image: vllm/vllm-openai-rocm:nightly-3f0a91bb96f8d72e0498b95c166e817deae14d62
   model: MiniMaxAI/MiniMax-M2.5
   model-prefix: minimaxm2.5
   runner: mi355x-disagg
@@ -1603,7 +1605,6 @@ minimaxm2.5-fp8-mi355x-vllm-disagg:
           dp-attn: false
           additional-settings:
           - "PREFILL_NODES=1"
-          - "VLLM_MORIIO_CONNECTOR_READ_MODE=1"
         decode:
           num-worker: 2
           tp: 8
@@ -1624,7 +1625,6 @@ minimaxm2.5-fp8-mi355x-vllm-disagg:
           dp-attn: false
           additional-settings:
           - "PREFILL_NODES=1"
-          - "VLLM_MORIIO_CONNECTOR_READ_MODE=1"
         decode:
           num-worker: 2
           tp: 8

benchmarks/multi_node/amd_utils/job.slurm

@@ -314,7 +314,7 @@ SANITIZED_USER=$(echo "$USER_NAME" | tr -c 'a-zA-Z0-9_.-' '_')
 export DOCKER_CONT_NAME="container_${ENGINE}_${SANITIZED_USER}_${MODEL_NAME}_${SLURM_JOB_ID}"
 
 # vLLM external router container
-VLLM_ROUTER_IMAGE="${VLLM_ROUTER_IMAGE:-vllm/vllm-router:nightly-20260511-e667ebb}"
+VLLM_ROUTER_IMAGE="${VLLM_ROUTER_IMAGE:-vllm/vllm-router:nightly-20260603-e667ebb}"
 ROUTER_CONT_NAME="router_vllm_${SANITIZED_USER}_${SLURM_JOB_ID}"
 export RUN_FILE_FULL="$WS_PATH/${RUN_FILE}"
 
@@ -399,7 +399,6 @@ if [[ "$ENGINE" == "vllm-disagg" ]]; then
         -e UCX_LOG_LEVEL=warn
         -e HSA_ENABLE_SDMA=1
         -e PROXY_STREAM_IDLE_TIMEOUT=\${PROXY_STREAM_IDLE_TIMEOUT:-300}
-        -e VLLM_MORIIO_CONNECTOR_READ_MODE=\${VLLM_MORIIO_CONNECTOR_READ_MODE:-1}
         -e PYTHONPYCACHEPREFIX=/tmp/pycache
     )
 else

benchmarks/multi_node/amd_utils/models_vllm.yaml

@@ -26,15 +26,15 @@ amd-Llama-3.3-70B-Instruct-FP8-KV:
 
 Kimi-K2.5-MXFP4:
   prefill_flags: "--tensor-parallel-size 8 --compilation-config '{\"cudagraph_mode\":\"PIECEWISE\"}' --no-enable-prefix-caching --block-size 1 --gpu-memory-utilization 0.90 --mm-encoder-tp-mode data"
-  decode_flags: "--tensor-parallel-size 8 --enable-expert-parallel --all2all-backend mori --compilation-config '{\"cudagraph_mode\":\"PIECEWISE\"}' --no-enable-prefix-caching --block-size 1 --gpu-memory-utilization 0.90 --mm-encoder-tp-mode data"
+  decode_flags: "--tensor-parallel-size 8 --enable-expert-parallel --all2all-backend mori_low_latency --compilation-config '{\"cudagraph_mode\":\"PIECEWISE\"}' --no-enable-prefix-caching --block-size 1 --gpu-memory-utilization 0.90 --mm-encoder-tp-mode data"
   env: "VLLM_USE_V1=1 VLLM_ROCM_USE_AITER=1 VLLM_ROCM_USE_AITER_PAGED_ATTN=0 VLLM_ROCM_USE_AITER_RMSNORM=1 VLLM_USE_AITER_TRITON_SILU_MUL=0 VLLM_ENGINE_READY_TIMEOUT_S=3600"
   hf_dir: "models--amd--Kimi-K2.5-MXFP4"
 
 MiniMax-M2.5:
   # AITER fused-MoE kernel fmoe_bf16_blockscaleFp8_g1u1_vs_silu_32x384 for gfx950 writes OOB when run with MiniMax's shapes at M=8K(=num batched tokens), crashing vllm during AITER warmup.
   # Set token budget to 4k to avoid using that shape, instead of disabling AITER_MOE.
-  prefill_flags: "--max-num-batched-tokens 4K --tensor-parallel-size 8 --enable-expert-parallel --all2all-backend mori --no-enable-prefix-caching --gpu-memory-utilization 0.95 --block-size 32"
-  decode_flags: "--max-num-batched-tokens 4K --tensor-parallel-size 8 --enable-expert-parallel --all2all-backend mori --no-enable-prefix-caching --gpu-memory-utilization 0.95 --block-size 32"
+  prefill_flags: "--max-num-batched-tokens 4K --tensor-parallel-size 8 --enable-expert-parallel --all2all-backend mori_low_latency --no-enable-prefix-caching --gpu-memory-utilization 0.95 --block-size 32"
+  decode_flags: "--max-num-batched-tokens 4K --tensor-parallel-size 8 --enable-expert-parallel --all2all-backend mori_low_latency --no-enable-prefix-caching --gpu-memory-utilization 0.95 --block-size 32"
   env: "VLLM_USE_V1=1 VLLM_ROCM_USE_AITER=1 VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 VLLM_ENGINE_READY_TIMEOUT_S=3600 VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=1"
   hf_dir: "models--MiniMaxAI--MiniMax-M2.5"
 

benchmarks/multi_node/amd_utils/server_vllm.sh

@@ -256,7 +256,7 @@ if [ "$NODE_RANK" -eq 0 ]; then
         --served-model-name ${SERVED_MODEL} \
         --port $SERVER_PORT \
         --trust-remote-code \
-        --kv-transfer-config '{\"kv_connector\": \"MoRIIOConnector\", \"kv_role\": \"kv_producer\", \"kv_connector_extra_config\": {\"proxy_ip\": \"${NODE0_ADDR}\", \"proxy_ping_port\": \"${PROXY_PING_PORT}\", \"http_port\": \"${SERVER_PORT}\"}}' \
+        --kv-transfer-config '{\"kv_connector\": \"MoRIIOConnector\", \"kv_role\": \"kv_producer\", \"kv_connector_extra_config\": {\"proxy_ip\": \"${NODE0_ADDR}\", \"proxy_ping_port\": \"${PROXY_PING_PORT}\", \"http_port\": \"${SERVER_PORT}\", \"read_mode\": true}}' \
         ${PREFILL_SERVER_CONFIG}"
 
     if [[ "$DRY_RUN" -eq 1 ]]; then
@@ -422,7 +422,7 @@ elif [ "$NODE_RANK" -gt 0 ] && [ "$NODE_RANK" -lt "$xP" ]; then
         --served-model-name ${SERVED_MODEL} \
         --port $SERVER_PORT \
         --trust-remote-code \
-        --kv-transfer-config '{\"kv_connector\": \"MoRIIOConnector\", \"kv_role\": \"kv_producer\", \"kv_connector_extra_config\": {\"proxy_ip\": \"${NODE0_ADDR}\", \"proxy_ping_port\": \"${PROXY_PING_PORT}\", \"http_port\": \"${SERVER_PORT}\"}}' \
+        --kv-transfer-config '{\"kv_connector\": \"MoRIIOConnector\", \"kv_role\": \"kv_producer\", \"kv_connector_extra_config\": {\"proxy_ip\": \"${NODE0_ADDR}\", \"proxy_ping_port\": \"${PROXY_PING_PORT}\", \"http_port\": \"${SERVER_PORT}\", \"read_mode\": true}}' \
         ${PREFILL_SERVER_CONFIG}"
 
     if [[ "$DRY_RUN" -eq 1 ]]; then
@@ -478,7 +478,7 @@ else
         --served-model-name ${SERVED_MODEL} \
         --port $SERVER_PORT \
         --trust-remote-code \
-        --kv-transfer-config '{\"kv_connector\": \"MoRIIOConnector\", \"kv_role\": \"kv_consumer\", \"kv_connector_extra_config\": {\"proxy_ip\": \"${NODE0_ADDR}\", \"proxy_ping_port\": \"${PROXY_PING_PORT}\", \"http_port\": \"${SERVER_PORT}\"}}' \
+        --kv-transfer-config '{\"kv_connector\": \"MoRIIOConnector\", \"kv_role\": \"kv_consumer\", \"kv_connector_extra_config\": {\"proxy_ip\": \"${NODE0_ADDR}\", \"proxy_ping_port\": \"${PROXY_PING_PORT}\", \"http_port\": \"${SERVER_PORT}\", \"read_mode\": true}}' \
         ${DECODE_SERVER_CONFIG}"
 
     if [[ "$DRY_RUN" -eq 1 ]]; then