[AMD][DI][CI] Add GLM-5.2 MXFP4 wide-EP16 2P1D nightly recipes (#34476)

scripts/ci/slurm/nightly-configs.yaml

@@ -604,3 +604,35 @@ glm52-fp4-mi355x-dp8ep8-mtp-sglang:
       search-space:
         - conc-list: [1, 8, 16, 32, 64, 128, 256]
           config_file: scripts/ci/slurm/recipes/mi355x-fp4/glm52/1k1k/1p1d-dp8ep8-mtp.yaml
+
+glm52-fp4-mi355x-ep16-sglang:
+  model: amd/GLM-5.2-MXFP4
+  model-prefix: glm52
+  model_path: /it-share/model_coverage/models--amd--GLM-5.1-MXFP4
+  runner: mi355x
+  precision: fp4
+  framework: sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+    - isl: 1024
+      osl: 1024
+      search-space:
+        - conc-list: [1, 8, 16, 32, 64, 128, 256]
+          config_file: scripts/ci/slurm/recipes/mi355x-fp4/glm52/1k1k/2p1d-ep16.yaml
+
+glm52-fp4-mi355x-ep16-mtp-sglang:
+  model: amd/GLM-5.2-MXFP4
+  model-prefix: glm52
+  model_path: /it-share/model_coverage/models--amd--GLM-5.1-MXFP4
+  runner: mi355x
+  precision: fp4
+  framework: sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+    - isl: 1024
+      osl: 1024
+      search-space:
+        - conc-list: [1, 8, 16, 32, 64, 128, 256]
+          config_file: scripts/ci/slurm/recipes/mi355x-fp4/glm52/1k1k/2p1d-ep16-mtp.yaml

scripts/ci/slurm/recipes/mi355x-fp4/glm52/1k1k/2p1d-ep16-mtp.yaml

@@ -0,0 +1,90 @@
+# MI355X GLM-5.2 (MXFP4) 4-node 2P1D disaggregation recipe -- narrow-prefill EP8 + MTP
+# + wide-decode EP16 (mirrors the DSV4-Pro / Kimi Oren config: wide EP only helps decode).
+#
+# Two prefill engines (EP8, one node each) + one decode engine (EP16) spanning 2
+# nodes, 4 nodes total. Prefill EP8 keeps MoE all-to-all INTRA-node (XGMI); decode
+# gets wide EP16 across nodes over mori. KV (prefill TP8 -> decode TP16) over mori.
+#
+# GLM-specific bits vs the Kimi EP16 recipe: DSA attention is auto-selected for
+# GlmMoeDsaForCausalLM (no explicit --attention-backend), GLM parsers (glm45),
+# and --disable-shared-experts-fusion (GLM has a shared expert).
+
+resources:
+  prefill_workers: 2
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+    decode:
+      tensor-parallel-size: 16
+      expert-parallel-size: 16
+      data-parallel-size: 16
+
+model:
+  env:
+    SGLANG_USE_AITER: 1
+  server_args:
+    - --reasoning-parser
+    - glm45
+    - --tool-call-parser
+    - glm45
+    - --disable-shared-experts-fusion
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.17-rocm720-mi35x-20260810
+  # DSA attention auto-selected for GlmMoeDsaForCausalLM (no --attention-backend).
+  ib_devices: rdma0,rdma1,rdma2,rdma3,rdma4,rdma5,rdma6,rdma7
+  moe_a2a_backend: mori
+  kv_transfer_backend: mori
+  dist_socket_ifname: eno0
+  rocm700a: 0
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 131072
+  swa_full_tokens_ratio: 0.1
+  wide_ep:
+    kv_cache_dtype: fp8_e4m3
+    prefill_mem_fraction_static: 0.85
+    decode_mem_fraction_static: 0.85
+    prefill_chunked_prefill_size: 131072
+    prefill_max_running_requests: 1024
+    decode_max_running_requests: 1024
+    common_extra_flags: "--moe-dense-tp-size 1 --enable-dp-lm-head --decode-log-interval 100 --watchdog-timeout 3600 --load-balance-method round_robin --dist-timeout 3600"
+    prefill_extra_flags: "--context-length 9217 --max-total-tokens 262144"
+    decode_extra_flags: "--disable-cuda-graph --prefill-round-robin-balance"
+    prefill_extra_env:
+      MORI_MAX_DISPATCH_TOKENS_PREFILL: 8192
+      MORI_MAX_DISPATCH_TOKENS_DECODE: 256
+      SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK: 16384
+    decode_extra_env:
+      MORI_MAX_DISPATCH_TOKENS_DECODE: 512
+      MORI_MOE_MAX_INPUT_TOKENS_DECODE: 2048
+      SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK: 1024
+
+# MTP / EAGLE speculative decoding (built-in NextN head from the base model).
+# Applied to both prefill and decode; no external draft checkpoint.
+mtp:
+  enabled: true
+  num_steps: 3
+  eagle_topk: 1
+  num_draft_tokens: 4
+
+bench:
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4
+  random_range_ratio: 1.0
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319
+    threshold: 0.91

scripts/ci/slurm/recipes/mi355x-fp4/glm52/1k1k/2p1d-ep16.yaml

@@ -0,0 +1,82 @@
+# MI355X GLM-5.2 (MXFP4) 4-node 2P1D disaggregation recipe -- narrow-prefill EP8
+# + wide-decode EP16 (mirrors the DSV4-Pro / Kimi Oren config: wide EP only helps decode).
+#
+# Two prefill engines (EP8, one node each) + one decode engine (EP16) spanning 2
+# nodes, 4 nodes total. Prefill EP8 keeps MoE all-to-all INTRA-node (XGMI); decode
+# gets wide EP16 across nodes over mori. KV (prefill TP8 -> decode TP16) over mori.
+#
+# GLM-specific bits vs the Kimi EP16 recipe: DSA attention is auto-selected for
+# GlmMoeDsaForCausalLM (no explicit --attention-backend), GLM parsers (glm45),
+# and --disable-shared-experts-fusion (GLM has a shared expert).
+
+resources:
+  prefill_workers: 2
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+    decode:
+      tensor-parallel-size: 16
+      expert-parallel-size: 16
+      data-parallel-size: 16
+
+model:
+  env:
+    SGLANG_USE_AITER: 1
+  server_args:
+    - --reasoning-parser
+    - glm45
+    - --tool-call-parser
+    - glm45
+    - --disable-shared-experts-fusion
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.17-rocm720-mi35x-20260810
+  # DSA attention auto-selected for GlmMoeDsaForCausalLM (no --attention-backend).
+  ib_devices: rdma0,rdma1,rdma2,rdma3,rdma4,rdma5,rdma6,rdma7
+  moe_a2a_backend: mori
+  kv_transfer_backend: mori
+  dist_socket_ifname: eno0
+  rocm700a: 0
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 131072
+  swa_full_tokens_ratio: 0.1
+  wide_ep:
+    kv_cache_dtype: fp8_e4m3
+    prefill_mem_fraction_static: 0.8
+    decode_mem_fraction_static: 0.85
+    prefill_chunked_prefill_size: 131072
+    prefill_max_running_requests: 1024
+    decode_max_running_requests: 1024
+    common_extra_flags: "--moe-dense-tp-size 1 --enable-dp-lm-head --decode-log-interval 100 --watchdog-timeout 3600 --load-balance-method round_robin --dist-timeout 3600"
+    prefill_extra_flags: "--context-length 9217 --max-total-tokens 262144"
+    decode_extra_flags: "--disable-cuda-graph --prefill-round-robin-balance"
+    prefill_extra_env:
+      MORI_MAX_DISPATCH_TOKENS_PREFILL: 8192
+      MORI_MAX_DISPATCH_TOKENS_DECODE: 256
+      SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK: 16384
+    decode_extra_env:
+      MORI_MAX_DISPATCH_TOKENS_DECODE: 64
+      MORI_MOE_MAX_INPUT_TOKENS_DECODE: 332
+      SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK: 128
+
+bench:
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4
+  random_range_ratio: 1.0
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319
+    threshold: 0.91