[Klaud Cold] Add MI355X config: qwen3.5-fp4-sglang-mtp (#1445)

* Add MI355X config: qwen3.5-fp4-sglang-mtp

Adds the qwen3.5-fp4-mi355x-sglang-mtp recipe using the latest
sglang-rocm nightly v0.5.12-rocm720-mi35x-20260517 (matches PRs
#1440-1444 which bump other mi355x recipes to the same image).

Replaces the stale #1078 which targeted v0.5.10rc0-rocm700-mi35x-20260417
and was tagged broken; that branch reverted unrelated merged changes
when rebased.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* chore: fill in pr-link for #1445 in perf-changelog

* fix(qwen3.5-fp4-mi355x-sglang-mtp): use scenarios/fixed-seq-len schema

The old 'seq-len-configs' field was renamed to 'scenarios.fixed-seq-len'
in the master config schema. Pydantic validation failed setup with
'Field required (scenarios)' + 'Extra inputs are not permitted (seq-len-configs)'.

* Add launch script benchmarks/single_node/qwen3.5_fp4_mi355x_mtp.sh

Carried over from the original #1078 branch — was missed in the
recipe-only port that opened #1445.

* Re-trigger sweep (abandon in-flight run with transient infra failures)

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: claude-fix-bot <claude-fix-bot@local>

Author: 3 people

.github/configs/amd-master.yaml

@@ -367,6 +367,27 @@ qwen3.5-fp4-mi355x-atom:
       - { tp: 2, conc-start: 4, conc-end: 256 }
       - { tp: 4, conc-start: 4, conc-end: 16 }
 
+qwen3.5-fp4-mi355x-sglang-mtp:
+  image: lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260517
+  model: amd/Qwen3.5-397B-A17B-MXFP4
+  model-prefix: qwen3.5
+  runner: mi355x
+  precision: fp4
+  framework: sglang
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 2, conc-start: 4, conc-end: 256, spec-decoding: mtp }
+      - { tp: 4, conc-start: 4, conc-end: 16, spec-decoding: mtp }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 2, conc-start: 4, conc-end: 256, spec-decoding: mtp }
+      - { tp: 4, conc-start: 4, conc-end: 16, spec-decoding: mtp }
+
 qwen3.5-fp8-mi300x-sglang:
   image: lmsysorg/sglang:v0.5.12-rocm720-mi30x
   model: Qwen/Qwen3.5-397B-A17B-FP8

benchmarks/single_node/qwen3.5_fp4_mi355x_mtp.sh

@@ -0,0 +1,73 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+hf download "$MODEL"
+
+export SGLANG_USE_AITER=1
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+MEM_FRAC_STATIC=${MEM_FRAC_STATIC:-0.8}
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+fi
+
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+set -x
+python3 -m sglang.launch_server --model-path=$MODEL --trust-remote-code \
+--host=0.0.0.0 --port=$PORT \
+--tensor-parallel-size=$TP \
+--attention-backend aiter \
+--mem-fraction-static $MEM_FRAC_STATIC \
+--model-loader-extra-config '{"enable_multithread_load": true}' \
+--watchdog-timeout 1200  \
+--disable-radix-cache \
+--speculative-algorithm EAGLE \
+--speculative-num-steps 3 \
+--speculative-eagle-topk 1 \
+--speculative-num-draft-tokens 4 \
+> $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+# Wait for server to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID" --sleep-interval 60
+
+run_benchmark_serving \
+    --model "$MODEL" \
+    --port "$PORT" \
+    --backend vllm \
+    --input-len "$ISL" \
+    --output-len "$OSL" \
+    --random-range-ratio "$RANDOM_RANGE_RATIO" \
+    --num-prompts "$((CONC * 10))" \
+    --max-concurrency "$CONC" \
+    --result-filename "$RESULT_FILENAME" \
+    --result-dir /workspace/
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x

perf-changelog.yaml

@@ -2895,3 +2895,9 @@
     - "Upgrade GB200 DSV4 MTP2 1P/1D DEP8 MegaMOE mid-curve recipe to vLLM v0.21.0"
     - "Extend mid-curve concurrency sweep to also benchmark conc=256, conc=512, and conc=1024"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1390
+
+- config-keys:
+    - qwen3.5-fp4-mi355x-sglang-mtp
+  description:
+    - "Add MI355X config: qwen3.5-fp4-sglang-mtp using lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260517"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1445