diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index f8cc486b2..30c657da1 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -9428,3 +9428,213 @@ glm5-fp4-gb300-dynamo-sglang:
           tp: 4
           ep: 1
           dp-attn: false
+
+glm5-fp8-gb300-dynamo-sglang:
+  image: lmsysorg/sglang:v0.5.11-cu130
+  model: zai-org/GLM-5-FP8
+  model-prefix: glm5
+  runner: gb300-nv
+  precision: fp8
+  framework: dynamo-sglang
+  multinode: true
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    # ---------- 8k1k high-throughput (wide-EP decode) ----------
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - conc-list: [2800]
+        prefill:
+          num-worker: 14
+          tp: 4
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_0.yaml"
+        decode:
+          num-worker: 1
+          tp: 16
+          ep: 16
+          dp-attn: true
+      - conc-list: [1700]
+        prefill:
+          num-worker: 12
+          tp: 4
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_1.yaml"
+        decode:
+          num-worker: 1
+          tp: 24
+          ep: 24
+          dp-attn: true
+      - conc-list: [1300]
+        prefill:
+          num-worker: 10
+          tp: 4
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_2.yaml"
+        decode:
+          num-worker: 1
+          tp: 32
+          ep: 32
+          dp-attn: true
+      - conc-list: [900]
+        prefill:
+          num-worker: 8
+          tp: 4
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_3.yaml"
+        decode:
+          num-worker: 1
+          tp: 40
+          ep: 40
+          dp-attn: true
+    # ---------- 8k1k low-latency (per-node TP=4 decode workers) ----------
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - conc-list: [150]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml"
+        decode:
+          num-worker: 9
+          tp: 4
+          ep: 1
+          dp-attn: false
+      - conc-list: [128, 64, 32]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_1.yaml"
+        decode:
+          num-worker: 17
+          tp: 4
+          ep: 1
+          dp-attn: false
+      - conc-list: [24]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_2.yaml"
+        decode:
+          num-worker: 17
+          tp: 4
+          ep: 1
+          dp-attn: false
+    # ---------- 1k1k high-throughput (wide-EP decode) ----------
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - conc-list: [8192]
+        prefill:
+          num-worker: 12
+          tp: 4
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_0.yaml"
+        decode:
+          num-worker: 1
+          tp: 24
+          ep: 24
+          dp-attn: true
+      - conc-list: [7500]
+        prefill:
+          num-worker: 10
+          tp: 4
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_1.yaml"
+        decode:
+          num-worker: 1
+          tp: 32
+          ep: 32
+          dp-attn: true
+      - conc-list: [7300]
+        prefill:
+          num-worker: 8
+          tp: 4
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_2.yaml"
+        decode:
+          num-worker: 1
+          tp: 40
+          ep: 40
+          dp-attn: true
+      - conc-list: [6500]
+        prefill:
+          num-worker: 6
+          tp: 4
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_3.yaml"
+        decode:
+          num-worker: 1
+          tp: 48
+          ep: 48
+          dp-attn: true
+      - conc-list: [5700]
+        prefill:
+          num-worker: 4
+          tp: 4
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_4.yaml"
+        decode:
+          num-worker: 1
+          tp: 56
+          ep: 56
+          dp-attn: true
+    # ---------- 1k1k low-latency (per-node TP=4 decode workers) ----------
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - conc-list: [512, 256, 128, 64]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_0.yaml"
+        decode:
+          num-worker: 17
+          tp: 4
+          ep: 1
+          dp-attn: false
+      - conc-list: [32]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_1.yaml"
+        decode:
+          num-worker: 17
+          tp: 4
+          ep: 1
+          dp-attn: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_0.yaml
new file mode 100644
index 000000000..33da57e94
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_0.yaml
@@ -0,0 +1,148 @@
+name: gb300-fp8-glm5_1k1k_hightpt_0
+
+model:
+  path: glm-5-fp8
+  container: "lmsysorg/sglang:v0.5.11-cu130"
+  precision: fp8
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 12
+  prefill_workers: 12
+  decode_nodes: 6
+  decode_workers: 1
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 9
+dynamo:
+  version: 1.1.0
+
+backend:
+  prefill_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+
+  decode_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+      # DeepEP per-rank dispatch buffer; must be >= ceil(cuda_graph_max_bs / dp_size).
+      # Default 128 overflows with large DP + batch (e.g. 4096/24 ~= 171 > 128). Limit 1024.
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '512'
+
+  sglang_config:
+    prefill:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: nixl
+
+        # Size limits
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      mem-fraction-static: 0.7
+      context-length: 9600
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 8192
+
+        # Parallelism
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 1
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      load-balance-method: total_tokens
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+      moe-runner-backend: flashinfer_trtllm
+
+        # Other flags
+      enable-flashinfer-allreduce-fusion: true
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+
+    decode:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: nixl
+
+        # Memory and token limits
+      mem-fraction-static: 0.8
+      context-length: 9600
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+        # moe-runner-backend: "cutedsl"
+
+        # Detokenizer
+      skip-tokenizer-init: true
+      stream-interval: 30
+
+        # Other flags
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+      tensor-parallel-size: 24
+      expert-parallel-size: 24
+      data-parallel-size: 24
+      enable-dp-lm-head: true
+      enable-dp-attention: true
+      moe-dense-tp-size: 1
+      ep-num-redundant-experts: 32
+      ep-dispatch-algorithm: static
+      moe-a2a-backend: deepep
+      deepep-mode: low_latency
+      deepep-config: /configs/deepep_config.json
+      max-running-requests: 8192
+      cuda-graph-max-bs: 512
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+benchmark:
+  type: sa-bench
+  req_rate: inf
+  isl: 1024
+  osl: 1024
+  concurrencies: '8192'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_1.yaml
new file mode 100644
index 000000000..03cb7e671
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_1.yaml
@@ -0,0 +1,148 @@
+name: gb300-fp8-glm5_1k1k_hightpt_1
+
+model:
+  path: glm-5-fp8
+  container: "lmsysorg/sglang:v0.5.11-cu130"
+  precision: fp8
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 10
+  prefill_workers: 10
+  decode_nodes: 8
+  decode_workers: 1
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 9
+dynamo:
+  version: 1.1.0
+
+backend:
+  prefill_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+
+  decode_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+      # DeepEP per-rank dispatch buffer; must be >= ceil(cuda_graph_max_bs / dp_size).
+      # Default 128 overflows with large DP + batch (e.g. 4096/24 ~= 171 > 128). Limit 1024.
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '512'
+
+  sglang_config:
+    prefill:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: nixl
+
+        # Size limits
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      mem-fraction-static: 0.7
+      context-length: 9600
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 8192
+
+        # Parallelism
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 1
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      load-balance-method: total_tokens
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+      moe-runner-backend: flashinfer_trtllm
+
+        # Other flags
+      enable-flashinfer-allreduce-fusion: true
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+
+    decode:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: nixl
+
+        # Memory and token limits
+      mem-fraction-static: 0.8
+      context-length: 9600
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+        # moe-runner-backend: "cutedsl"
+
+        # Detokenizer
+      skip-tokenizer-init: true
+      stream-interval: 30
+
+        # Other flags
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+      tensor-parallel-size: 32
+      expert-parallel-size: 32
+      data-parallel-size: 32
+      enable-dp-lm-head: true
+      enable-dp-attention: true
+      moe-dense-tp-size: 1
+      ep-num-redundant-experts: 32
+      ep-dispatch-algorithm: static
+      moe-a2a-backend: deepep
+      deepep-mode: low_latency
+      deepep-config: /configs/deepep_config.json
+      max-running-requests: 8192
+      cuda-graph-max-bs: 256
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+benchmark:
+  type: sa-bench
+  req_rate: inf
+  isl: 1024
+  osl: 1024
+  concurrencies: '7500'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_2.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_2.yaml
new file mode 100644
index 000000000..c06206c81
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_2.yaml
@@ -0,0 +1,148 @@
+name: gb300-fp8-glm5_1k1k_hightpt_2
+
+model:
+  path: glm-5-fp8
+  container: "lmsysorg/sglang:v0.5.11-cu130"
+  precision: fp8
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 8
+  prefill_workers: 8
+  decode_nodes: 10
+  decode_workers: 1
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 9
+dynamo:
+  version: 1.1.0
+
+backend:
+  prefill_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+
+  decode_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+      # DeepEP per-rank dispatch buffer; must be >= ceil(cuda_graph_max_bs / dp_size).
+      # Default 128 overflows with large DP + batch (e.g. 4096/24 ~= 171 > 128). Limit 1024.
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '512'
+
+  sglang_config:
+    prefill:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: nixl
+
+        # Size limits
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      mem-fraction-static: 0.7
+      context-length: 9600
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 8192
+
+        # Parallelism
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 1
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      load-balance-method: total_tokens
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+      moe-runner-backend: flashinfer_trtllm
+
+        # Other flags
+      enable-flashinfer-allreduce-fusion: true
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+
+    decode:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: nixl
+
+        # Memory and token limits
+      mem-fraction-static: 0.8
+      context-length: 9600
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+        # moe-runner-backend: "cutedsl"
+
+        # Detokenizer
+      skip-tokenizer-init: true
+      stream-interval: 30
+
+        # Other flags
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+      tensor-parallel-size: 40
+      expert-parallel-size: 40
+      data-parallel-size: 40
+      enable-dp-lm-head: true
+      enable-dp-attention: true
+      moe-dense-tp-size: 1
+      ep-num-redundant-experts: 24
+      ep-dispatch-algorithm: static
+      moe-a2a-backend: deepep
+      deepep-mode: low_latency
+      deepep-config: /configs/deepep_config.json
+      max-running-requests: 7200
+      cuda-graph-max-bs: 180
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+benchmark:
+  type: sa-bench
+  req_rate: inf
+  isl: 1024
+  osl: 1024
+  concurrencies: '7300'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_3.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_3.yaml
new file mode 100644
index 000000000..951772479
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_3.yaml
@@ -0,0 +1,148 @@
+name: gb300-fp8-glm5_1k1k_hightpt_3
+
+model:
+  path: glm-5-fp8
+  container: "lmsysorg/sglang:v0.5.11-cu130"
+  precision: fp8
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 6
+  prefill_workers: 6
+  decode_nodes: 12
+  decode_workers: 1
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 9
+dynamo:
+  version: 1.1.0
+
+backend:
+  prefill_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+
+  decode_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+      # DeepEP per-rank dispatch buffer; must be >= ceil(cuda_graph_max_bs / dp_size).
+      # Default 128 overflows with large DP + batch (e.g. 4096/24 ~= 171 > 128). Limit 1024.
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '512'
+
+  sglang_config:
+    prefill:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: nixl
+
+        # Size limits
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      mem-fraction-static: 0.7
+      context-length: 9600
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 8192
+
+        # Parallelism
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 1
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      load-balance-method: total_tokens
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+      moe-runner-backend: flashinfer_trtllm
+
+        # Other flags
+      enable-flashinfer-allreduce-fusion: true
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+
+    decode:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: nixl
+
+        # Memory and token limits
+      mem-fraction-static: 0.8
+      context-length: 9600
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+        # moe-runner-backend: "cutedsl"
+
+        # Detokenizer
+      skip-tokenizer-init: true
+      stream-interval: 30
+
+        # Other flags
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+      tensor-parallel-size: 48
+      expert-parallel-size: 48
+      data-parallel-size: 48
+      enable-dp-lm-head: true
+      enable-dp-attention: true
+      moe-dense-tp-size: 1
+      ep-num-redundant-experts: 32
+      ep-dispatch-algorithm: static
+      moe-a2a-backend: deepep
+      deepep-mode: low_latency
+      deepep-config: /configs/deepep_config.json
+      max-running-requests: 6144
+      cuda-graph-max-bs: 128
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+benchmark:
+  type: sa-bench
+  req_rate: inf
+  isl: 1024
+  osl: 1024
+  concurrencies: '6500'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_4.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_4.yaml
new file mode 100644
index 000000000..9a1f320a5
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_4.yaml
@@ -0,0 +1,148 @@
+name: gb300-fp8-glm5_1k1k_hightpt_4
+
+model:
+  path: glm-5-fp8
+  container: "lmsysorg/sglang:v0.5.11-cu130"
+  precision: fp8
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 4
+  prefill_workers: 4
+  decode_nodes: 14
+  decode_workers: 1
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 9
+dynamo:
+  version: 1.1.0
+
+backend:
+  prefill_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+
+  decode_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+      # DeepEP per-rank dispatch buffer; must be >= ceil(cuda_graph_max_bs / dp_size).
+      # Default 128 overflows with large DP + batch (e.g. 4096/24 ~= 171 > 128). Limit 1024.
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '512'
+
+  sglang_config:
+    prefill:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: nixl
+
+        # Size limits
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      mem-fraction-static: 0.7
+      context-length: 9600
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 8192
+
+        # Parallelism
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 1
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      load-balance-method: total_tokens
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+      moe-runner-backend: flashinfer_trtllm
+
+        # Other flags
+      enable-flashinfer-allreduce-fusion: true
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+
+    decode:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: nixl
+
+        # Memory and token limits
+      mem-fraction-static: 0.8
+      context-length: 9600
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+        # moe-runner-backend: "cutedsl"
+
+        # Detokenizer
+      skip-tokenizer-init: true
+      stream-interval: 30
+
+        # Other flags
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+      tensor-parallel-size: 56
+      expert-parallel-size: 56
+      data-parallel-size: 56
+      enable-dp-lm-head: true
+      enable-dp-attention: true
+      moe-dense-tp-size: 1
+      ep-num-redundant-experts: 24
+      ep-dispatch-algorithm: static
+      moe-a2a-backend: deepep
+      deepep-mode: low_latency
+      deepep-config: /configs/deepep_config.json
+      max-running-requests: 5600
+      cuda-graph-max-bs: 100
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+benchmark:
+  type: sa-bench
+  req_rate: inf
+  isl: 1024
+  osl: 1024
+  concurrencies: '5700'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_0.yaml
new file mode 100644
index 000000000..3ace5647c
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_0.yaml
@@ -0,0 +1,140 @@
+name: gb300-fp8-glm5_1k1k_lowlat_0
+
+model:
+  path: glm-5-fp8
+  container: "lmsysorg/sglang:v0.5.11-cu130"
+  precision: fp8
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 1
+  decode_nodes: 17
+  decode_workers: 17
+frontend:
+  type: dynamo
+dynamo:
+  version: 1.1.0
+
+backend:
+  prefill_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+
+  decode_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+      # DeepEP per-rank dispatch buffer; must be >= ceil(cuda_graph_max_bs / dp_size).
+      # Default 128 overflows with large DP + batch (e.g. 4096/24 ~= 171 > 128). Limit 1024.
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '512'
+
+  sglang_config:
+    prefill:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: nixl
+
+        # Size limits
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      mem-fraction-static: 0.7
+      context-length: 9600
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 8192
+
+        # Parallelism
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 1
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      load-balance-method: total_tokens
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+      moe-runner-backend: flashinfer_trtllm
+
+        # Other flags
+      enable-flashinfer-allreduce-fusion: true
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+
+    decode:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: nixl
+
+        # Memory and token limits
+      mem-fraction-static: 0.8
+      context-length: 9600
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+        # moe-runner-backend: "cutedsl"
+
+        # Detokenizer
+      skip-tokenizer-init: true
+      stream-interval: 30
+
+        # Other flags
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+      tensor-parallel-size: 4
+      expert-parallel-size: 1
+      data-parallel-size: 1
+      enable-flashinfer-allreduce-fusion: true
+      moe-runner-backend: flashinfer_trtllm
+      max-running-requests: 32
+      cuda-graph-max-bs: 32
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+benchmark:
+  type: sa-bench
+  req_rate: inf
+  isl: 1024
+  osl: 1024
+  concurrencies: 512x256x128x64
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_1.yaml
new file mode 100644
index 000000000..965b6f148
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_1.yaml
@@ -0,0 +1,140 @@
+name: gb300-fp8-glm5_1k1k_lowlat_1
+
+model:
+  path: glm-5-fp8
+  container: "lmsysorg/sglang:v0.5.11-cu130"
+  precision: fp8
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 1
+  decode_nodes: 17
+  decode_workers: 17
+frontend:
+  type: dynamo
+dynamo:
+  version: 1.1.0
+
+backend:
+  prefill_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+
+  decode_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+      # DeepEP per-rank dispatch buffer; must be >= ceil(cuda_graph_max_bs / dp_size).
+      # Default 128 overflows with large DP + batch (e.g. 4096/24 ~= 171 > 128). Limit 1024.
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '512'
+
+  sglang_config:
+    prefill:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: nixl
+
+        # Size limits
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      mem-fraction-static: 0.7
+      context-length: 9600
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 8192
+
+        # Parallelism
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 1
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      load-balance-method: total_tokens
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+      moe-runner-backend: flashinfer_trtllm
+
+        # Other flags
+      enable-flashinfer-allreduce-fusion: true
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+
+    decode:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: nixl
+
+        # Memory and token limits
+      mem-fraction-static: 0.8
+      context-length: 9600
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+        # moe-runner-backend: "cutedsl"
+
+        # Detokenizer
+      skip-tokenizer-init: true
+      stream-interval: 30
+
+        # Other flags
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+      tensor-parallel-size: 4
+      expert-parallel-size: 1
+      data-parallel-size: 1
+      enable-flashinfer-allreduce-fusion: true
+      moe-runner-backend: flashinfer_trtllm
+      max-running-requests: 1
+      cuda-graph-max-bs: 1
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+benchmark:
+  type: sa-bench
+  req_rate: inf
+  isl: 1024
+  osl: 1024
+  concurrencies: '32'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_0.yaml
new file mode 100644
index 000000000..150e62233
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_0.yaml
@@ -0,0 +1,148 @@
+name: gb300-fp8-glm5_8k1k_hightpt_0
+
+model:
+  path: glm-5-fp8
+  container: "lmsysorg/sglang:v0.5.11-cu130"
+  precision: fp8
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 14
+  prefill_workers: 14
+  decode_nodes: 4
+  decode_workers: 1
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 9
+dynamo:
+  version: 1.1.0
+
+backend:
+  prefill_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+
+  decode_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+      # DeepEP per-rank dispatch buffer; must be >= ceil(cuda_graph_max_bs / dp_size).
+      # Default 128 overflows with large DP + batch (e.g. 4096/24 ~= 171 > 128). Limit 1024.
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '512'
+
+  sglang_config:
+    prefill:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: nixl
+
+        # Size limits
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      mem-fraction-static: 0.7
+      context-length: 9600
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 8192
+
+        # Parallelism
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 1
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      load-balance-method: total_tokens
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+      moe-runner-backend: flashinfer_trtllm
+
+        # Other flags
+      enable-flashinfer-allreduce-fusion: true
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+
+    decode:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: nixl
+
+        # Memory and token limits
+      mem-fraction-static: 0.8
+      context-length: 9600
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+        # moe-runner-backend: "cutedsl"
+
+        # Detokenizer
+      skip-tokenizer-init: true
+      stream-interval: 30
+
+        # Other flags
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+      tensor-parallel-size: 16
+      expert-parallel-size: 16
+      data-parallel-size: 16
+      enable-dp-lm-head: true
+      enable-dp-attention: true
+      moe-dense-tp-size: 1
+      ep-num-redundant-experts: 32
+      ep-dispatch-algorithm: static
+      moe-a2a-backend: deepep
+      deepep-mode: low_latency
+      deepep-config: /configs/deepep_config.json
+      max-running-requests: 2800
+      cuda-graph-max-bs: 175
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+benchmark:
+  type: sa-bench
+  req_rate: inf
+  isl: 8192
+  osl: 1024
+  concurrencies: '2800'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_1.yaml
new file mode 100644
index 000000000..6393069c8
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_1.yaml
@@ -0,0 +1,148 @@
+name: gb300-fp8-glm5_8k1k_hightpt_1
+
+model:
+  path: glm-5-fp8
+  container: "lmsysorg/sglang:v0.5.11-cu130"
+  precision: fp8
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 12
+  prefill_workers: 12
+  decode_nodes: 6
+  decode_workers: 1
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 9
+dynamo:
+  version: 1.1.0
+
+backend:
+  prefill_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+
+  decode_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+      # DeepEP per-rank dispatch buffer; must be >= ceil(cuda_graph_max_bs / dp_size).
+      # Default 128 overflows with large DP + batch (e.g. 4096/24 ~= 171 > 128). Limit 1024.
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '512'
+
+  sglang_config:
+    prefill:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: nixl
+
+        # Size limits
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      mem-fraction-static: 0.7
+      context-length: 9600
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 8192
+
+        # Parallelism
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 1
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      load-balance-method: total_tokens
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+      moe-runner-backend: flashinfer_trtllm
+
+        # Other flags
+      enable-flashinfer-allreduce-fusion: true
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+
+    decode:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: nixl
+
+        # Memory and token limits
+      mem-fraction-static: 0.8
+      context-length: 9600
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+        # moe-runner-backend: "cutedsl"
+
+        # Detokenizer
+      skip-tokenizer-init: true
+      stream-interval: 30
+
+        # Other flags
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+      tensor-parallel-size: 24
+      expert-parallel-size: 24
+      data-parallel-size: 24
+      enable-dp-lm-head: true
+      enable-dp-attention: true
+      moe-dense-tp-size: 1
+      ep-num-redundant-experts: 32
+      ep-dispatch-algorithm: static
+      moe-a2a-backend: deepep
+      deepep-mode: low_latency
+      deepep-config: /configs/deepep_config.json
+      max-running-requests: 1680
+      cuda-graph-max-bs: 70
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+benchmark:
+  type: sa-bench
+  req_rate: inf
+  isl: 8192
+  osl: 1024
+  concurrencies: '1700'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_2.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_2.yaml
new file mode 100644
index 000000000..56b11ed8a
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_2.yaml
@@ -0,0 +1,148 @@
+name: gb300-fp8-glm5_8k1k_hightpt_2
+
+model:
+  path: glm-5-fp8
+  container: "lmsysorg/sglang:v0.5.11-cu130"
+  precision: fp8
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 10
+  prefill_workers: 10
+  decode_nodes: 8
+  decode_workers: 1
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 9
+dynamo:
+  version: 1.1.0
+
+backend:
+  prefill_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+
+  decode_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+      # DeepEP per-rank dispatch buffer; must be >= ceil(cuda_graph_max_bs / dp_size).
+      # Default 128 overflows with large DP + batch (e.g. 4096/24 ~= 171 > 128). Limit 1024.
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '512'
+
+  sglang_config:
+    prefill:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: nixl
+
+        # Size limits
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      mem-fraction-static: 0.7
+      context-length: 9600
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 8192
+
+        # Parallelism
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 1
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      load-balance-method: total_tokens
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+      moe-runner-backend: flashinfer_trtllm
+
+        # Other flags
+      enable-flashinfer-allreduce-fusion: true
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+
+    decode:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: nixl
+
+        # Memory and token limits
+      mem-fraction-static: 0.8
+      context-length: 9600
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+        # moe-runner-backend: "cutedsl"
+
+        # Detokenizer
+      skip-tokenizer-init: true
+      stream-interval: 30
+
+        # Other flags
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+      tensor-parallel-size: 32
+      expert-parallel-size: 32
+      data-parallel-size: 32
+      enable-dp-lm-head: true
+      enable-dp-attention: true
+      moe-dense-tp-size: 1
+      ep-num-redundant-experts: 32
+      ep-dispatch-algorithm: static
+      moe-a2a-backend: deepep
+      deepep-mode: low_latency
+      deepep-config: /configs/deepep_config.json
+      max-running-requests: 1280
+      cuda-graph-max-bs: 40
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+benchmark:
+  type: sa-bench
+  req_rate: inf
+  isl: 8192
+  osl: 1024
+  concurrencies: '1300'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_3.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_3.yaml
new file mode 100644
index 000000000..13fb0f326
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_3.yaml
@@ -0,0 +1,148 @@
+name: gb300-fp8-glm5_8k1k_hightpt_3
+
+model:
+  path: glm-5-fp8
+  container: "lmsysorg/sglang:v0.5.11-cu130"
+  precision: fp8
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 8
+  prefill_workers: 8
+  decode_nodes: 10
+  decode_workers: 1
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 9
+dynamo:
+  version: 1.1.0
+
+backend:
+  prefill_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+
+  decode_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+      # DeepEP per-rank dispatch buffer; must be >= ceil(cuda_graph_max_bs / dp_size).
+      # Default 128 overflows with large DP + batch (e.g. 4096/24 ~= 171 > 128). Limit 1024.
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '512'
+
+  sglang_config:
+    prefill:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: nixl
+
+        # Size limits
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      mem-fraction-static: 0.7
+      context-length: 9600
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 8192
+
+        # Parallelism
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 1
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      load-balance-method: total_tokens
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+      moe-runner-backend: flashinfer_trtllm
+
+        # Other flags
+      enable-flashinfer-allreduce-fusion: true
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+
+    decode:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: nixl
+
+        # Memory and token limits
+      mem-fraction-static: 0.8
+      context-length: 9600
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+        # moe-runner-backend: "cutedsl"
+
+        # Detokenizer
+      skip-tokenizer-init: true
+      stream-interval: 30
+
+        # Other flags
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+      tensor-parallel-size: 40
+      expert-parallel-size: 40
+      data-parallel-size: 40
+      enable-dp-lm-head: true
+      enable-dp-attention: true
+      moe-dense-tp-size: 1
+      ep-num-redundant-experts: 24
+      ep-dispatch-algorithm: static
+      moe-a2a-backend: deepep
+      deepep-mode: low_latency
+      deepep-config: /configs/deepep_config.json
+      max-running-requests: 880
+      cuda-graph-max-bs: 22
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+benchmark:
+  type: sa-bench
+  req_rate: inf
+  isl: 8192
+  osl: 1024
+  concurrencies: '900'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml
new file mode 100644
index 000000000..8065160bd
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml
@@ -0,0 +1,140 @@
+name: gb300-fp8-glm5_8k1k_lowlat_0
+
+model:
+  path: glm-5-fp8
+  container: "lmsysorg/sglang:v0.5.11-cu130"
+  precision: fp8
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 1
+  decode_nodes: 9
+  decode_workers: 9
+frontend:
+  type: dynamo
+dynamo:
+  version: 1.1.0
+
+backend:
+  prefill_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+
+  decode_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+      # DeepEP per-rank dispatch buffer; must be >= ceil(cuda_graph_max_bs / dp_size).
+      # Default 128 overflows with large DP + batch (e.g. 4096/24 ~= 171 > 128). Limit 1024.
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '512'
+
+  sglang_config:
+    prefill:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: nixl
+
+        # Size limits
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      mem-fraction-static: 0.7
+      context-length: 9600
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 8192
+
+        # Parallelism
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 1
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      load-balance-method: total_tokens
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+      moe-runner-backend: flashinfer_trtllm
+
+        # Other flags
+      enable-flashinfer-allreduce-fusion: true
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+
+    decode:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: nixl
+
+        # Memory and token limits
+      mem-fraction-static: 0.8
+      context-length: 9600
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+        # moe-runner-backend: "cutedsl"
+
+        # Detokenizer
+      skip-tokenizer-init: true
+      stream-interval: 30
+
+        # Other flags
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+      tensor-parallel-size: 4
+      expert-parallel-size: 1
+      data-parallel-size: 1
+      enable-flashinfer-allreduce-fusion: true
+      moe-runner-backend: flashinfer_trtllm
+      max-running-requests: 15
+      cuda-graph-max-bs: 15
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+benchmark:
+  type: sa-bench
+  req_rate: inf
+  isl: 8192
+  osl: 1024
+  concurrencies: '150'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_1.yaml
new file mode 100644
index 000000000..33f0324b6
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_1.yaml
@@ -0,0 +1,140 @@
+name: gb300-fp8-glm5_8k1k_lowlat_1
+
+model:
+  path: glm-5-fp8
+  container: "lmsysorg/sglang:v0.5.11-cu130"
+  precision: fp8
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 1
+  decode_nodes: 17
+  decode_workers: 17
+frontend:
+  type: dynamo
+dynamo:
+  version: 1.1.0
+
+backend:
+  prefill_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+
+  decode_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+      # DeepEP per-rank dispatch buffer; must be >= ceil(cuda_graph_max_bs / dp_size).
+      # Default 128 overflows with large DP + batch (e.g. 4096/24 ~= 171 > 128). Limit 1024.
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '512'
+
+  sglang_config:
+    prefill:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: nixl
+
+        # Size limits
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      mem-fraction-static: 0.7
+      context-length: 9600
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 8192
+
+        # Parallelism
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 1
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      load-balance-method: total_tokens
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+      moe-runner-backend: flashinfer_trtllm
+
+        # Other flags
+      enable-flashinfer-allreduce-fusion: true
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+
+    decode:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: nixl
+
+        # Memory and token limits
+      mem-fraction-static: 0.8
+      context-length: 9600
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+        # moe-runner-backend: "cutedsl"
+
+        # Detokenizer
+      skip-tokenizer-init: true
+      stream-interval: 30
+
+        # Other flags
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+      tensor-parallel-size: 4
+      expert-parallel-size: 1
+      data-parallel-size: 1
+      enable-flashinfer-allreduce-fusion: true
+      moe-runner-backend: flashinfer_trtllm
+      max-running-requests: 8
+      cuda-graph-max-bs: 8
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+benchmark:
+  type: sa-bench
+  req_rate: inf
+  isl: 8192
+  osl: 1024
+  concurrencies: 128x64x32
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_2.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_2.yaml
new file mode 100644
index 000000000..64d4c701a
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_2.yaml
@@ -0,0 +1,140 @@
+name: gb300-fp8-glm5_8k1k_lowlat_2
+
+model:
+  path: glm-5-fp8
+  container: "lmsysorg/sglang:v0.5.11-cu130"
+  precision: fp8
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 1
+  decode_nodes: 17
+  decode_workers: 17
+frontend:
+  type: dynamo
+dynamo:
+  version: 1.1.0
+
+backend:
+  prefill_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+
+  decode_environment:
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+    PYTHONUNBUFFERED: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    MC_TE_METRIC: 'true'
+    MC_FORCE_MNNVL: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    DYN_REQUEST_PLANE: nats
+      # DeepEP per-rank dispatch buffer; must be >= ceil(cuda_graph_max_bs / dp_size).
+      # Default 128 overflows with large DP + batch (e.g. 4096/24 ~= 171 > 128). Limit 1024.
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '512'
+
+  sglang_config:
+    prefill:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: nixl
+
+        # Size limits
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      mem-fraction-static: 0.7
+      context-length: 9600
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 8192
+
+        # Parallelism
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 1
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      load-balance-method: total_tokens
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+      moe-runner-backend: flashinfer_trtllm
+
+        # Other flags
+      enable-flashinfer-allreduce-fusion: true
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+
+    decode:
+        # Model configuration
+      served-model-name: GLM-5-FP8
+      trust-remote-code: true
+
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+
+        # Disaggregation mode
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: nixl
+
+        # Memory and token limits
+      mem-fraction-static: 0.8
+      context-length: 9600
+
+        # Backend
+      nsa-decode-backend: trtllm
+      nsa-prefill-backend: trtllm
+        # moe-runner-backend: "cutedsl"
+
+        # Detokenizer
+      skip-tokenizer-init: true
+      stream-interval: 30
+
+        # Other flags
+      disable-radix-cache: true
+      weight-loader-prefetch-checkpoints: true
+      model-loader-extra-config: '{"enable_multithread_load": true}'
+      tensor-parallel-size: 4
+      expert-parallel-size: 1
+      data-parallel-size: 1
+      enable-flashinfer-allreduce-fusion: true
+      moe-runner-backend: flashinfer_trtllm
+      max-running-requests: 1
+      cuda-graph-max-bs: 1
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+benchmark:
+  type: sa-bench
+  req_rate: inf
+  isl: 8192
+  osl: 1024
+  concurrencies: '24'
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 935cded22..1b194b52f 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3200,3 +3200,11 @@
     - "Bump image to lmsysorg/sglang-rocm:v0.5.12.post1-rocm720-mi35x-20260523, 1P1D TP8/EP1, dp-attn false, conc [8..512]"
     - "MoRI conn.py overlay (48e459bd) via job.slurm; launcher qwen3.5_fp4_mi355x_sglang-disagg.sh"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1579
+
+- config-keys:
+    - glm5-fp8-gb300-dynamo-sglang
+  description:
+    - "Add GLM-5 FP8 GB300 Dynamo SGLang disaggregated multi-node coverage using lmsysorg/sglang:v0.5.11-cu130"
+    - "1k1k and 8k1k STP hightpt and lowlat srt-slurm recipes under benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp8/ (resolved from upstream srt-slurm PR #160 via srtctl resolve-override)"
+    - "Wire glm5/fp8 model + dynamo-sglang framework branches into runners/launch_gb300-nv.sh with SA upstream defaults (SLURM_PARTITION=batch_1, SLURM_ACCOUNT=benchmark, SQUASH_FILE under /home/sa-shared/gharunners/squash/)"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1557
diff --git a/runners/launch_gb300-cw.sh b/runners/launch_gb300-cw.sh
index 25e7f4db5..676362c01 100644
--- a/runners/launch_gb300-cw.sh
+++ b/runners/launch_gb300-cw.sh
@@ -25,8 +25,20 @@ if [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" ]]; then
         echo "Unsupported framework on gb300-cw for dsv4/fp4: $FRAMEWORK. Currently supported: dynamo-sglang, dynamo-vllm"
         exit 1
     fi
+elif [[ $MODEL_PREFIX == "glm5" && $PRECISION == "fp8" ]]; then
+    export MODEL_PATH="/mnt/vast/models/GLM-5-FP8"
+
+    if [[ $FRAMEWORK == "dynamo-sglang" ]]; then
+        SRT_SLURM_RECIPES_REPO="https://github.com/NVIDIA/srt-slurm.git"
+        SRT_SLURM_RECIPES_REF="sa-submission-q2-2026"
+        SRT_RECIPE_SRC="$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5"
+        SRT_RECIPE_DST="recipes/sglang/glm5"
+    else
+        echo "Unsupported framework on gb300-cw for glm5/fp8: $FRAMEWORK. Currently supported: dynamo-sglang"
+        exit 1
+    fi
 else
-    echo "Unsupported model prefix/precision combination on gb300-cw: $MODEL_PREFIX/$PRECISION. Currently supported: dsv4/fp4"
+    echo "Unsupported model prefix/precision combination on gb300-cw: $MODEL_PREFIX/$PRECISION. Currently supported: dsv4/fp4, glm5/fp8"
     exit 1
 fi
 
@@ -183,6 +195,8 @@ model_paths:
   # is not a local model path and is not defined in srtslurm.yaml
   # model_paths".
   deepseek-v4-pro: "${MODEL_PATH}"
+  # GLM-5 FP8 sglang recipes use `model.path: glm-5-fp8`.
+  glm-5-fp8: "${MODEL_PATH}"
 containers:
   dynamo-trtllm: ${SQUASH_FILE}
   dynamo-sglang: ${SQUASH_FILE}
diff --git a/runners/launch_gb300-nv.sh b/runners/launch_gb300-nv.sh
index 5248e63ed..56ee8033d 100644
--- a/runners/launch_gb300-nv.sh
+++ b/runners/launch_gb300-nv.sh
@@ -24,8 +24,11 @@ elif [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" ]]; then
 elif [[ $MODEL_PREFIX == "glm5" && $PRECISION == "fp4" ]]; then
     export MODEL_PATH=/scratch/models/GLM-5-NVFP4
     export SRT_SLURM_MODEL_PREFIX="glm-5-fp4"
+elif [[ $MODEL_PREFIX == "glm5" && $PRECISION == "fp8" ]]; then
+    export MODEL_PATH=/scratch/models/GLM-5-FP8
+    export SRT_SLURM_MODEL_PREFIX="glm-5-fp8"
 else
-    echo "Unsupported model: $MODEL_PREFIX-$PRECISION. Supported models are: dsr1-fp4, dsr1-fp8, dsv4-fp4, glm5-fp4"
+    echo "Unsupported model: $MODEL_PREFIX-$PRECISION. Supported models are: dsr1-fp4, dsr1-fp8, dsv4-fp4, glm5-fp4, glm5-fp8"
     exit 1
 fi
 
@@ -72,11 +75,6 @@ if [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
     mkdir -p recipes/vllm/deepseek-v4
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4" recipes/vllm/deepseek-v4
 elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "glm5" ]]; then
-    # Hand-rolled GLM-5 sglang recipes (one flat yaml per topology,
-    # derived from upstream srt-slurm PR #152's combined glm5.yaml).
-    # Directory layout mirrors SemiAnalysisAI/InferenceX PR #1372:
-    #   sglang/glm5/<hw>-<precision>/<seq>/disagg/stp/<seq>_stp_<variant>_<idx>.yaml
-    # so multiple precisions (fp4/fp8) and hardware targets can coexist.
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR"
     git checkout sa-submission-q2-2026