From 8a1e08b1468878f5a2443202c7e72ff96b58c8a5 Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Mon, 18 May 2026 02:05:52 -0400
Subject: [PATCH 1/3] Add glm5-fp8-mi300x-sglang (off + mtp) recipes

New family on MI300X using lmsysorg/sglang:v0.5.12-rocm720-mi30x.
TP=8, conc 4..64, 1k1k + 8k1k. Launch scripts follow the
qwen3.5-fp8-mi300x SGLang recipe (aiter attention + AMD allreduce
fusion) and add GLM-5 parsers (glm47 tool calls, glm45 reasoning).

MTP variant adds --speculative-algorithm EAGLE plus the standard
EAGLE knobs (num-steps 3, eagle-topk 1, num-draft-tokens 4) and
--use-chat-template on the bench client per AGENTS.md.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .github/configs/amd-master.yaml               | 38 +++++++++
 benchmarks/single_node/glm5_fp8_mi300x.sh     | 76 +++++++++++++++++
 benchmarks/single_node/glm5_fp8_mi300x_mtp.sh | 84 +++++++++++++++++++
 perf-changelog.yaml                           |  7 ++
 4 files changed, 205 insertions(+)
 create mode 100755 benchmarks/single_node/glm5_fp8_mi300x.sh
 create mode 100755 benchmarks/single_node/glm5_fp8_mi300x_mtp.sh

diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index 83542d07a..809d48473 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -1933,3 +1933,41 @@ glm5-fp8-mi325x-sglang-mtp:
       osl: 1024
       search-space:
       - { tp: 8, ep: 1, conc-start: 4, conc-end: 64, spec-decoding: mtp }
+
+glm5-fp8-mi300x-sglang:
+  image: lmsysorg/sglang:v0.5.12-rocm720-mi30x
+  model: zai-org/GLM-5-FP8
+  model-prefix: glm5
+  runner: mi300x
+  precision: fp8
+  framework: sglang
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 4, conc-end: 64 }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 4, conc-end: 64 }
+
+glm5-fp8-mi300x-sglang-mtp:
+  image: lmsysorg/sglang:v0.5.12-rocm720-mi30x
+  model: zai-org/GLM-5-FP8
+  model-prefix: glm5
+  runner: mi300x
+  precision: fp8
+  framework: sglang
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 8, ep: 1, conc-start: 4, conc-end: 64, spec-decoding: mtp }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 8, ep: 1, conc-start: 4, conc-end: 64, spec-decoding: mtp }
diff --git a/benchmarks/single_node/glm5_fp8_mi300x.sh b/benchmarks/single_node/glm5_fp8_mi300x.sh
new file mode 100755
index 000000000..01205192a
--- /dev/null
+++ b/benchmarks/single_node/glm5_fp8_mi300x.sh
@@ -0,0 +1,76 @@
+#!/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 $SLURMD_NODENAME"
+fi
+
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+CONTEXT_LENGTH=$((ISL + OSL + 20))
+MAX_PREFILL_TOKENS=32768
+
+EVAL_CONTEXT_ARGS=""
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN"
+else EVAL_CONTEXT_ARGS="--context-length $CONTEXT_LENGTH"
+fi
+
+start_gpu_monitor
+
+# Follows the qwen3.5-fp8-mi300x recipe (AMD Andy's LinkedIn-recommended
+# args for sglang ROCm on mi300x) with GLM-5-specific parsers.
+python3 -m sglang.launch_server \
+    --attention-backend aiter \
+    --model-path $MODEL \
+    --host=0.0.0.0 \
+    --port $PORT \
+    --tensor-parallel-size $TP \
+    --data-parallel-size 1 \
+    --trust-remote-code \
+    --tool-call-parser glm47 \
+    --reasoning-parser glm45 \
+    --tokenizer-worker-num 6 \
+    --enable-aiter-allreduce-fusion \
+    --cuda-graph-max-bs $CONC \
+    --disable-radix-cache \
+    --max-prefill-tokens $MAX_PREFILL_TOKENS \
+    --scheduler-recv-interval 30 \
+    --mem-fraction-static 0.75 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+run_benchmark_serving \
+    --model "$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 "$RESULT_FILENAME" \
+    --result-dir /workspace/
+
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+stop_gpu_monitor
+set +x
diff --git a/benchmarks/single_node/glm5_fp8_mi300x_mtp.sh b/benchmarks/single_node/glm5_fp8_mi300x_mtp.sh
new file mode 100755
index 000000000..f9b7dae93
--- /dev/null
+++ b/benchmarks/single_node/glm5_fp8_mi300x_mtp.sh
@@ -0,0 +1,84 @@
+#!/usr/bin/env bash
+
+# GLM-5 FP8 on MI300X with EAGLE / MTP speculative decoding.
+# Mirrors glm5_fp8_mi300x.sh and adds the speculative-* flags.
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME \
+    EP_SIZE
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+CONTEXT_LENGTH=$((ISL + OSL + 20))
+MAX_PREFILL_TOKENS=32768
+
+EVAL_CONTEXT_ARGS=""
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN"
+else EVAL_CONTEXT_ARGS="--context-length $CONTEXT_LENGTH"
+fi
+
+start_gpu_monitor
+
+python3 -m sglang.launch_server \
+    --attention-backend aiter \
+    --model-path $MODEL \
+    --host=0.0.0.0 \
+    --port $PORT \
+    --tensor-parallel-size $TP \
+    --ep-size $EP_SIZE \
+    --trust-remote-code \
+    --tool-call-parser glm47 \
+    --reasoning-parser glm45 \
+    --tokenizer-worker-num 6 \
+    --enable-aiter-allreduce-fusion \
+    --cuda-graph-max-bs $CONC \
+    --disable-radix-cache \
+    --max-prefill-tokens $MAX_PREFILL_TOKENS \
+    --scheduler-recv-interval 30 \
+    --mem-fraction-static 0.75 \
+    --speculative-algorithm EAGLE \
+    --speculative-num-steps 3 \
+    --speculative-eagle-topk 1 \
+    --speculative-num-draft-tokens 4 \
+    $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+run_benchmark_serving \
+    --model "$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 "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --use-chat-template
+
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+stop_gpu_monitor
+set +x
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 9c4894c47..2435877f5 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3050,3 +3050,10 @@
   description:
     - "Update SGLang image from v0.5.11-cu130 (5d old) to v0.5.12-cu130"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1475
+
+- config-keys:
+    - glm5-fp8-mi300x-sglang
+    - glm5-fp8-mi300x-sglang-mtp
+  description:
+    - "Add GLM-5 FP8 SGLang ROCm recipes (off + MTP/EAGLE) for MI300X on lmsysorg/sglang:v0.5.12-rocm720-mi30x"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX

From eb4caeafcdaf5844e005da47f267267dfb88dd6e Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Mon, 18 May 2026 02:06:05 -0400
Subject: [PATCH 2/3] chore: fill pr-link for #1486

---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 2435877f5..ca194e9aa 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3056,4 +3056,4 @@
     - glm5-fp8-mi300x-sglang-mtp
   description:
     - "Add GLM-5 FP8 SGLang ROCm recipes (off + MTP/EAGLE) for MI300X on lmsysorg/sglang:v0.5.12-rocm720-mi30x"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1486

From 8fe48f720174cebabe22eafbafe5f07aa874a77c Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Tue, 19 May 2026 18:51:50 -0400
Subject: [PATCH 3/3] Switch GLM-5 mi300x launch args to tilelang NSA +
 fp8_e4m3 KV cache

Drops --attention-backend aiter / --enable-aiter-allreduce-fusion in
favor of the recipe suggested in sglang issue #25672 comment 4485916205:
  --mem-fraction-static 0.80
  --model-loader-extra-config '{"enable_multithread_load": true, "num_threads": 8}'
  --nsa-prefill-backend tilelang --nsa-decode-backend tilelang
  --kv-cache-dtype fp8_e4m3

Same fix applied to glm5-fp8-mi325x in #1485; both recipes share the
aiter Triton MHA tl.arange power-of-2 crash on GLM-5. Applied to both
the off and MTP variants.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 benchmarks/single_node/glm5_fp8_mi300x.sh     | 13 ++++++++-----
 benchmarks/single_node/glm5_fp8_mi300x_mtp.sh | 11 ++++++++---
 2 files changed, 16 insertions(+), 8 deletions(-)

diff --git a/benchmarks/single_node/glm5_fp8_mi300x.sh b/benchmarks/single_node/glm5_fp8_mi300x.sh
index 01205192a..c1d24b76d 100755
--- a/benchmarks/single_node/glm5_fp8_mi300x.sh
+++ b/benchmarks/single_node/glm5_fp8_mi300x.sh
@@ -31,10 +31,9 @@ fi
 
 start_gpu_monitor
 
-# Follows the qwen3.5-fp8-mi300x recipe (AMD Andy's LinkedIn-recommended
-# args for sglang ROCm on mi300x) with GLM-5-specific parsers.
+# Launch args follow sglang issue #25672 comment 4485916205:
+# tilelang NSA backends + fp8_e4m3 KV cache + multithread model load.
 python3 -m sglang.launch_server \
-    --attention-backend aiter \
     --model-path $MODEL \
     --host=0.0.0.0 \
     --port $PORT \
@@ -44,12 +43,16 @@ python3 -m sglang.launch_server \
     --tool-call-parser glm47 \
     --reasoning-parser glm45 \
     --tokenizer-worker-num 6 \
-    --enable-aiter-allreduce-fusion \
     --cuda-graph-max-bs $CONC \
     --disable-radix-cache \
     --max-prefill-tokens $MAX_PREFILL_TOKENS \
     --scheduler-recv-interval 30 \
-    --mem-fraction-static 0.75 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
+    --mem-fraction-static 0.80 \
+    --model-loader-extra-config '{"enable_multithread_load": true, "num_threads": 8}' \
+    --nsa-prefill-backend tilelang \
+    --nsa-decode-backend tilelang \
+    --kv-cache-dtype fp8_e4m3 \
+    $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
 
 SERVER_PID=$!
 
diff --git a/benchmarks/single_node/glm5_fp8_mi300x_mtp.sh b/benchmarks/single_node/glm5_fp8_mi300x_mtp.sh
index f9b7dae93..f8f03c34c 100755
--- a/benchmarks/single_node/glm5_fp8_mi300x_mtp.sh
+++ b/benchmarks/single_node/glm5_fp8_mi300x_mtp.sh
@@ -35,8 +35,10 @@ fi
 
 start_gpu_monitor
 
+# Launch args follow sglang issue #25672 comment 4485916205:
+# tilelang NSA backends + fp8_e4m3 KV cache + multithread model load,
+# plus EAGLE/MTP speculative decoding.
 python3 -m sglang.launch_server \
-    --attention-backend aiter \
     --model-path $MODEL \
     --host=0.0.0.0 \
     --port $PORT \
@@ -46,12 +48,15 @@ python3 -m sglang.launch_server \
     --tool-call-parser glm47 \
     --reasoning-parser glm45 \
     --tokenizer-worker-num 6 \
-    --enable-aiter-allreduce-fusion \
     --cuda-graph-max-bs $CONC \
     --disable-radix-cache \
     --max-prefill-tokens $MAX_PREFILL_TOKENS \
     --scheduler-recv-interval 30 \
-    --mem-fraction-static 0.75 \
+    --mem-fraction-static 0.80 \
+    --model-loader-extra-config '{"enable_multithread_load": true, "num_threads": 8}' \
+    --nsa-prefill-backend tilelang \
+    --nsa-decode-backend tilelang \
+    --kv-cache-dtype fp8_e4m3 \
     --speculative-algorithm EAGLE \
     --speculative-num-steps 3 \
     --speculative-eagle-topk 1 \