From b5d87c772c007fc8781e0e8c46679d78303823fa Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 12 Aug 2026 12:06:44 -0500
Subject: [PATCH 1/9] Add MI300X GLM-5.2 FP8 AgentX discovery

---
 .../agentic/glm5.2_fp8_mi300x_mtp.sh          | 116 ++++++++++++++++++
 configs/amd-master.yaml                       |  17 +++
 perf-changelog.yaml                           |   8 ++
 3 files changed, 141 insertions(+)
 create mode 100755 benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh

diff --git a/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh b/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh
new file mode 100755
index 000000000..69c6bf1db
--- /dev/null
+++ b/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh
@@ -0,0 +1,116 @@
+#!/usr/bin/env bash
+set -eo pipefail
+set -x
+
+# GLM-5.2 FP8 on one 8xMI300X node with native EAGLE MTP.
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+export EVAL_FRAMEWORK="lm-eval"
+
+check_env_vars \
+    MODEL TP CONC EP_SIZE KV_OFFLOADING \
+    RESULT_DIR DURATION
+
+require_agentic_kv_offload_none
+
+if [[ -n "${SLURM_JOB_ID:-}" ]]; then
+    echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
+fi
+if [[ -n "${ROCR_VISIBLE_DEVICES:-}" ]]; then
+    export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
+fi
+
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
+rocm-smi || true
+amd-smi || true
+
+export WEKA_LOADER_OVERRIDE=semianalysis_cc_traces_weka_062126
+resolve_trace_source
+install_agentic_deps
+
+SERVER_LOG="$RESULT_DIR/server.log"
+mkdir -p "$RESULT_DIR"
+SERVER_PID=""
+
+cleanup_agentic_services() {
+    local exit_code=$?
+    trap - EXIT INT TERM
+    set +e
+    stop_background_process_tree "$SERVER_PID" "SGLang server" 60
+    exit "$exit_code"
+}
+trap cleanup_agentic_services EXIT
+trap 'exit 130' INT
+trap 'exit 143' TERM
+
+export PYTHONNOUSERSITE=1
+export AIPERF_HTTP_TCP_USER_TIMEOUT=900000
+export SGLANG_TIMEOUT_KEEP_ALIVE=900
+export SGLANG_DSA_FUSE_TOPK=false
+export SGLANG_OPT_USE_TOPK_V2=false
+
+# Throughput replays use the committed GLM-5.2 thinking-on K=3 golden AL;
+# evals retain real target-model verification.
+if [[ "${EVAL_ONLY:-false}" != "true" ]]; then
+    export SGLANG_SIMULATE_ACC_LEN=2.99
+    export SGLANG_SIMULATE_ACC_METHOD=match-expected
+    export SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token
+fi
+
+MAX_RUNNING_REQUESTS=$((2 * CONC))
+
+SGLANG_CMD=(
+    python3 -m sglang.launch_server
+    --model-path "$MODEL_PATH"
+    --served-model-name "$MODEL"
+    --host 0.0.0.0
+    --port "$PORT"
+    --trust-remote-code
+    --tp "$TP"
+    --ep-size "$EP_SIZE"
+    --dsa-prefill-backend tilelang
+    --dsa-decode-backend tilelang
+    --dsa-topk-backend torch
+    --kv-cache-dtype bfloat16
+    --tool-call-parser glm47
+    --reasoning-parser glm45
+    --context-length 1048576
+    --max-total-tokens 1048576
+    --chunked-prefill-size 131072
+    --mem-fraction-static 0.85
+    --max-running-requests "$MAX_RUNNING_REQUESTS"
+    --cuda-graph-max-bs "$MAX_RUNNING_REQUESTS"
+    --speculative-algorithm EAGLE
+    --speculative-num-steps 3
+    --speculative-eagle-topk 1
+    --speculative-num-draft-tokens 4
+    --watchdog-timeout 1800
+    --enable-metrics
+    --enable-cache-report
+)
+
+write_command "$RESULT_DIR/sglang_command.txt" "${SGLANG_CMD[@]}"
+"${SGLANG_CMD[@]}" > "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+if [[ "${EVAL_ONLY:-false}" == "true" ]]; then
+    export SWEBENCH_AGENT_STEP_LIMIT=150
+    run_eval --port "$PORT"
+else
+    # Aggregate serving exposes one logical SGLang Prometheus target.
+    export AIPERF_SERVER_METRICS_URLS="http://localhost:$PORT/metrics"
+    export AIPERF_REQUIRED_SERVER_METRIC_PREFIX="sglang:"
+    build_replay_cmd "$RESULT_DIR"
+    REPLAY_CMD+=" --benchmark-grace-period 1800"
+    run_agentic_replay_and_write_outputs "$RESULT_DIR"
+fi
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index c94029484..817646431 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1539,6 +1539,23 @@ glm5.2-fp8-mi325x-sglang-agentic-mtp:
       search-space:
       - { tp: 8, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 3, 4, 5, 6, 8] }
 
+# Dense discovery around the MI325X c3-c5 HBM knee. MI300X has 25% less HBM,
+# so sample every concurrency through c8 for both TP and expert-parallel MoE.
+glm5.2-fp8-mi300x-sglang-agentic-mtp:
+  image: lmsysorg/sglang:v0.5.16-rocm720-mi30x
+  model: zai-org/GLM-5.2-FP8
+  model-prefix: glm5.2
+  runner: cluster:mi300x-amds
+  precision: fp8
+  framework: sglang
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      - { tp: 8, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 3, 4, 5, 6, 7, 8] }
+      - { tp: 8, ep: 8, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 3, 4, 5, 6, 7, 8] }
+
 minimaxm3-fp8-mi325x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
   model: MiniMaxAI/MiniMax-M3-MXFP8
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 6082f2b86..e89c7fd77 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5918,3 +5918,11 @@
     - "Add a TEP2 arm (tp 2, ep 2) to the qwen3.5-fp4-b200-sglang-mtp 8k/1k sweep at concurrency 16, 32, and 64"
     - "Rides on the NVFP4-V2 checkpoint switch from #2205"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2550
+
+- config-keys:
+    - glm5.2-fp8-mi300x-sglang-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add MI300X GLM-5.2 FP8 AgentX discovery with native SGLang EAGLE MTP, golden AL 2.99, required SGLang metrics, and dense TP8/TEP8 c1-c8 resident sweeps."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/TBD

From 4bc7485b6357dc07826bc28bc5a1dceab4d5f52e Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 12 Aug 2026 12:07:11 -0500
Subject: [PATCH 2/9] Link MI300X GLM AgentX changelog

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

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index e89c7fd77..625dd99ce 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5925,4 +5925,4 @@
     - agentic-coding
   description:
     - "Add MI300X GLM-5.2 FP8 AgentX discovery with native SGLang EAGLE MTP, golden AL 2.99, required SGLang metrics, and dense TP8/TEP8 c1-c8 resident sweeps."
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/TBD
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2581

From 6551bbd2ed83cc2317f3012b0d4ccc0dab73a846 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 12 Aug 2026 13:31:41 -0500
Subject: [PATCH 3/9] Reserve MI300X DSA prefill headroom

---
 benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh b/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh
index 69c6bf1db..4a409fdc9 100755
--- a/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh
@@ -9,7 +9,7 @@ source "$(dirname "$0")/../../benchmark_lib.sh"
 export EVAL_FRAMEWORK="lm-eval"
 
 check_env_vars \
-    MODEL TP CONC EP_SIZE KV_OFFLOADING \
+    MODEL TP CONC EP_SIZE KV_OFFLOADING PORT EVAL_ONLY \
     RESULT_DIR DURATION
 
 require_agentic_kv_offload_none
@@ -85,7 +85,9 @@ SGLANG_CMD=(
     --context-length 1048576
     --max-total-tokens 1048576
     --chunked-prefill-size 131072
-    --mem-fraction-static 0.85
+    # Full 131072-token DSA prefills need transient workspace beyond the KV
+    # pool; 0.80 leaves enough headroom on 192 GB MI300X ranks.
+    --mem-fraction-static 0.80
     --max-running-requests "$MAX_RUNNING_REQUESTS"
     --cuda-graph-max-bs "$MAX_RUNNING_REQUESTS"
     --speculative-algorithm EAGLE

From 332490be13272282a32bba7639567f5aa3c3e66a Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 12 Aug 2026 14:56:22 -0500
Subject: [PATCH 4/9] Increase MI300X DSA prefill headroom

---
 benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh b/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh
index 4a409fdc9..037e922c9 100755
--- a/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh
@@ -86,8 +86,8 @@ SGLANG_CMD=(
     --max-total-tokens 1048576
     --chunked-prefill-size 131072
     # Full 131072-token DSA prefills need transient workspace beyond the KV
-    # pool; 0.80 leaves enough headroom on 192 GB MI300X ranks.
-    --mem-fraction-static 0.80
+    # pool; 0.75 leaves enough headroom on 192 GB MI300X ranks.
+    --mem-fraction-static 0.75
     --max-running-requests "$MAX_RUNNING_REQUESTS"
     --cuda-graph-max-bs "$MAX_RUNNING_REQUESTS"
     --speculative-algorithm EAGLE

From 5798048482d81f175e863fe4504a74abf6fe2e40 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 12 Aug 2026 17:37:37 -0500
Subject: [PATCH 5/9] Stabilize MI300X long-context prefills

---
 benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh b/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh
index 037e922c9..7f6f86a0c 100755
--- a/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh
@@ -86,8 +86,8 @@ SGLANG_CMD=(
     --max-total-tokens 1048576
     --chunked-prefill-size 131072
     # Full 131072-token DSA prefills need transient workspace beyond the KV
-    # pool; 0.75 leaves enough headroom on 192 GB MI300X ranks.
-    --mem-fraction-static 0.75
+    # pool; 0.70 leaves enough headroom on 192 GB MI300X ranks.
+    --mem-fraction-static 0.70
     --max-running-requests "$MAX_RUNNING_REQUESTS"
     --cuda-graph-max-bs "$MAX_RUNNING_REQUESTS"
     --speculative-algorithm EAGLE

From a67dd87393471c0ce4166c422aad38a6084895ce Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 12 Aug 2026 20:05:09 -0500
Subject: [PATCH 6/9] Test MI300X GLM HiCache at the c3 knee

---
 .../agentic/glm5.2_fp8_mi300x_mtp.sh          | 29 +++++++++++++++++--
 configs/amd-master.yaml                       |  3 +-
 2 files changed, 27 insertions(+), 5 deletions(-)

diff --git a/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh b/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh
index 7f6f86a0c..3f770c073 100755
--- a/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh
@@ -10,9 +10,7 @@ export EVAL_FRAMEWORK="lm-eval"
 
 check_env_vars \
     MODEL TP CONC EP_SIZE KV_OFFLOADING PORT EVAL_ONLY \
-    RESULT_DIR DURATION
-
-require_agentic_kv_offload_none
+    TOTAL_CPU_DRAM_GB RESULT_DIR DURATION
 
 if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
@@ -40,6 +38,30 @@ SERVER_LOG="$RESULT_DIR/server.log"
 mkdir -p "$RESULT_DIR"
 SERVER_PID=""
 
+CACHE_ARGS=()
+if require_agentic_kv_offload_backend hicache; then
+    # SGLang allocates one target and one much smaller EAGLE draft host pool
+    # per rank. Use the generated node-total DRAM budget, reserving one GB per
+    # rank for alignment; 64 GB per pool is enough to exercise eviction while
+    # preserving host headroom for model loading on MI300X nodes.
+    HICACHE_ALIGNMENT_RESERVE_GB=$TP
+    HICACHE_MAX_SIZE_GB=$(((TOTAL_CPU_DRAM_GB - HICACHE_ALIGNMENT_RESERVE_GB) / TP / 2))
+    HICACHE_SIZE_GB=64
+    if [ "$HICACHE_SIZE_GB" -gt "$HICACHE_MAX_SIZE_GB" ]; then
+        echo "Error: HiCache requires ${HICACHE_SIZE_GB} GB per rank pool but the configured limit is ${HICACHE_MAX_SIZE_GB} GB" >&2
+        exit 1
+    fi
+    echo "HiCache CPU pools: ${HICACHE_SIZE_GB} GB per rank pool across TP=${TP}, within ${TOTAL_CPU_DRAM_GB} GB node budget"
+    CACHE_ARGS=(
+        --page-size 64
+        --enable-hierarchical-cache
+        --hicache-size "$HICACHE_SIZE_GB"
+        --hicache-io-backend kernel
+        --hicache-mem-layout page_first
+        --hicache-write-policy write_through_selective
+    )
+fi
+
 cleanup_agentic_services() {
     local exit_code=$?
     trap - EXIT INT TERM
@@ -97,6 +119,7 @@ SGLANG_CMD=(
     --watchdog-timeout 1800
     --enable-metrics
     --enable-cache-report
+    "${CACHE_ARGS[@]}"
 )
 
 write_command "$RESULT_DIR/sglang_command.txt" "${SGLANG_CMD[@]}"
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 817646431..d0189c627 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1553,8 +1553,7 @@ glm5.2-fp8-mi300x-sglang-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 8, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 3, 4, 5, 6, 7, 8] }
-      - { tp: 8, ep: 8, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 3, 4, 5, 6, 7, 8] }
+      - { tp: 8, ep: 8, kv-offloading: dram, kv-offload-backend: { name: hicache }, spec-decoding: mtp, conc-list: [3] }
 
 minimaxm3-fp8-mi325x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914

From b8bb7af5855bd953205d84d5fa6ad2fba01c3f63 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 12 Aug 2026 20:31:58 -0500
Subject: [PATCH 7/9] Select MI300X GLM resident Pareto matrix

---
 .../agentic/glm5.2_fp8_mi300x_mtp.sh          | 29 ++-----------------
 configs/amd-master.yaml                       |  7 +++--
 perf-changelog.yaml                           |  2 +-
 3 files changed, 8 insertions(+), 30 deletions(-)

diff --git a/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh b/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh
index 3f770c073..7f6f86a0c 100755
--- a/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp8_mi300x_mtp.sh
@@ -10,7 +10,9 @@ export EVAL_FRAMEWORK="lm-eval"
 
 check_env_vars \
     MODEL TP CONC EP_SIZE KV_OFFLOADING PORT EVAL_ONLY \
-    TOTAL_CPU_DRAM_GB RESULT_DIR DURATION
+    RESULT_DIR DURATION
+
+require_agentic_kv_offload_none
 
 if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
@@ -38,30 +40,6 @@ SERVER_LOG="$RESULT_DIR/server.log"
 mkdir -p "$RESULT_DIR"
 SERVER_PID=""
 
-CACHE_ARGS=()
-if require_agentic_kv_offload_backend hicache; then
-    # SGLang allocates one target and one much smaller EAGLE draft host pool
-    # per rank. Use the generated node-total DRAM budget, reserving one GB per
-    # rank for alignment; 64 GB per pool is enough to exercise eviction while
-    # preserving host headroom for model loading on MI300X nodes.
-    HICACHE_ALIGNMENT_RESERVE_GB=$TP
-    HICACHE_MAX_SIZE_GB=$(((TOTAL_CPU_DRAM_GB - HICACHE_ALIGNMENT_RESERVE_GB) / TP / 2))
-    HICACHE_SIZE_GB=64
-    if [ "$HICACHE_SIZE_GB" -gt "$HICACHE_MAX_SIZE_GB" ]; then
-        echo "Error: HiCache requires ${HICACHE_SIZE_GB} GB per rank pool but the configured limit is ${HICACHE_MAX_SIZE_GB} GB" >&2
-        exit 1
-    fi
-    echo "HiCache CPU pools: ${HICACHE_SIZE_GB} GB per rank pool across TP=${TP}, within ${TOTAL_CPU_DRAM_GB} GB node budget"
-    CACHE_ARGS=(
-        --page-size 64
-        --enable-hierarchical-cache
-        --hicache-size "$HICACHE_SIZE_GB"
-        --hicache-io-backend kernel
-        --hicache-mem-layout page_first
-        --hicache-write-policy write_through_selective
-    )
-fi
-
 cleanup_agentic_services() {
     local exit_code=$?
     trap - EXIT INT TERM
@@ -119,7 +97,6 @@ SGLANG_CMD=(
     --watchdog-timeout 1800
     --enable-metrics
     --enable-cache-report
-    "${CACHE_ARGS[@]}"
 )
 
 write_command "$RESULT_DIR/sglang_command.txt" "${SGLANG_CMD[@]}"
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index d0189c627..5e8be4245 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1539,8 +1539,8 @@ glm5.2-fp8-mi325x-sglang-agentic-mtp:
       search-space:
       - { tp: 8, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 3, 4, 5, 6, 8] }
 
-# Dense discovery around the MI325X c3-c5 HBM knee. MI300X has 25% less HBM,
-# so sample every concurrency through c8 for both TP and expert-parallel MoE.
+# Fast discovery found a sharp quality cliff after c3. Keep TP c1-c3 plus TEP
+# c3, whose lower TTFT trades slightly lower input throughput than TP c3.
 glm5.2-fp8-mi300x-sglang-agentic-mtp:
   image: lmsysorg/sglang:v0.5.16-rocm720-mi30x
   model: zai-org/GLM-5.2-FP8
@@ -1553,7 +1553,8 @@ glm5.2-fp8-mi300x-sglang-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 8, ep: 8, kv-offloading: dram, kv-offload-backend: { name: hicache }, spec-decoding: mtp, conc-list: [3] }
+      - { tp: 8, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 3] }
+      - { tp: 8, ep: 8, kv-offloading: none, spec-decoding: mtp, conc-list: [3] }
 
 minimaxm3-fp8-mi325x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 625dd99ce..55e843cf4 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5924,5 +5924,5 @@
   scenario-type:
     - agentic-coding
   description:
-    - "Add MI300X GLM-5.2 FP8 AgentX discovery with native SGLang EAGLE MTP, golden AL 2.99, required SGLang metrics, and dense TP8/TEP8 c1-c8 resident sweeps."
+    - "Add MI300X GLM-5.2 FP8 AgentX with native SGLang EAGLE MTP and the verified resident Pareto matrix: TP8 c1-c3 plus TEP8 c3."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2581

From 6e265386646e0aeb3f4a3f7be42666665e17b3e5 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 13 Aug 2026 01:29:20 -0500
Subject: [PATCH 8/9] Drop unstable MI300X GLM c2 point

---
 configs/amd-master.yaml | 6 +++---
 perf-changelog.yaml     | 2 +-
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 5e8be4245..ad04d5453 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1539,8 +1539,8 @@ glm5.2-fp8-mi325x-sglang-agentic-mtp:
       search-space:
       - { tp: 8, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 3, 4, 5, 6, 8] }
 
-# Fast discovery found a sharp quality cliff after c3. Keep TP c1-c3 plus TEP
-# c3, whose lower TTFT trades slightly lower input throughput than TP c3.
+# Fast discovery found a sharp quality cliff after c3. Keep TP c1/c3 plus TEP
+# c3; TP c2 exhausted HBM in two identical official long-context warmups.
 glm5.2-fp8-mi300x-sglang-agentic-mtp:
   image: lmsysorg/sglang:v0.5.16-rocm720-mi30x
   model: zai-org/GLM-5.2-FP8
@@ -1553,7 +1553,7 @@ glm5.2-fp8-mi300x-sglang-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 8, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 3] }
+      - { tp: 8, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 3] }
       - { tp: 8, ep: 8, kv-offloading: none, spec-decoding: mtp, conc-list: [3] }
 
 minimaxm3-fp8-mi325x-vllm-agentic:
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 55e843cf4..a4e69d211 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5924,5 +5924,5 @@
   scenario-type:
     - agentic-coding
   description:
-    - "Add MI300X GLM-5.2 FP8 AgentX with native SGLang EAGLE MTP and the verified resident Pareto matrix: TP8 c1-c3 plus TEP8 c3."
+    - "Add MI300X GLM-5.2 FP8 AgentX with native SGLang EAGLE MTP and the verified resident Pareto matrix: TP8 c1/c3 plus TEP8 c3."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2581

From 7d22a04cb78d73116285396dee086cf23b49f113 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 13 Aug 2026 09:35:57 -0500
Subject: [PATCH 9/9] Drop unstable MI300X GLM c1 point

---
 configs/amd-master.yaml | 6 +++---
 perf-changelog.yaml     | 2 +-
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index ad04d5453..c43273973 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1539,8 +1539,8 @@ glm5.2-fp8-mi325x-sglang-agentic-mtp:
       search-space:
       - { tp: 8, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 3, 4, 5, 6, 8] }
 
-# Fast discovery found a sharp quality cliff after c3. Keep TP c1/c3 plus TEP
-# c3; TP c2 exhausted HBM in two identical official long-context warmups.
+# Fast discovery found a sharp quality cliff after c3. Keep the verified TP and
+# TEP c3 points; lower concurrency repeatedly failed quality or memory gates.
 glm5.2-fp8-mi300x-sglang-agentic-mtp:
   image: lmsysorg/sglang:v0.5.16-rocm720-mi30x
   model: zai-org/GLM-5.2-FP8
@@ -1553,7 +1553,7 @@ glm5.2-fp8-mi300x-sglang-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 8, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 3] }
+      - { tp: 8, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [3] }
       - { tp: 8, ep: 8, kv-offloading: none, spec-decoding: mtp, conc-list: [3] }
 
 minimaxm3-fp8-mi325x-vllm-agentic:
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index a4e69d211..76f745fe9 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5924,5 +5924,5 @@
   scenario-type:
     - agentic-coding
   description:
-    - "Add MI300X GLM-5.2 FP8 AgentX with native SGLang EAGLE MTP and the verified resident Pareto matrix: TP8 c1/c3 plus TEP8 c3."
+    - "Add MI300X GLM-5.2 FP8 AgentX with native SGLang EAGLE MTP and the verified resident Pareto matrix: TP8 c3 plus TEP8 c3."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2581