[AMD][DI][CI] 5/N Add DSV4 wide-EP16 4-node 2P1D nightly recipes (#31500)

Co-authored-by: Chen <bingxche@amd.com>

scripts/ci/slurm/nightly-configs.yaml

@@ -357,6 +357,149 @@ kimik26-fp8-mi355x-mtp-sglang:
         - conc-list: [1, 8, 16, 32, 64, 128, 256]
           config_file: scripts/ci/slurm/recipes/mi355x-fp8/kimik26/1k1k/1p1d-mtp.yaml
 
+# AMD 4-node disaggregation with narrow-prefill EP8 + WIDE-decode EP16 (Oren's
+# 2P1D config: two single-node prefill engines EP8 that the router fans across +
+# one decode engine EP16 spanning 2 nodes; 4 nodes total). Runs on the `mi355x`
+# amd-sglang cluster (bnxt RoCE), not spur: spur's ionic fabric could not cross-
+# rail the MORI MoE all-to-all, so EP16 was brought up and validated on mi355x
+# (job 13221, DSV4-Pro-FP4, GSM8K 0.927). Each recipe sets
+# runtime.moe_a2a_backend=mori + runtime.kv_transfer_backend=mori +
+# runtime.ib_devices=rdma0..7 + runtime.dist_socket_ifname=eno0; launch_mi355x.sh
+# derives nodes-per-engine = ceil(TP/8) (prefill 8->1, decode 16->2) and emits the
+# cross-node --nnodes/--node-rank/--dist-init-addr args for the decode engine.
+# DSV4-Pro MTP drops conc256 (SWA retract->get_cpu_copy NotImplementedError).
+dsv4flash-fp8-mi355x-ep16-sglang:
+  model: sgl-project/DeepSeek-V4-Flash-FP8
+  model-prefix: dsv4flash
+  model_path: /it-share/model_coverage/models--sgl-project--DeepSeek-V4-Flash-FP8
+  runner: mi355x
+  precision: fp8
+  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-fp8/dsv4flash/1k1k/2p1d-ep16.yaml
+
+dsv4flash-fp8-mi355x-ep16-mtp-sglang:
+  model: sgl-project/DeepSeek-V4-Flash-FP8
+  model-prefix: dsv4flash
+  model_path: /it-share/model_coverage/models--sgl-project--DeepSeek-V4-Flash-FP8
+  runner: mi355x
+  precision: fp8
+  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-fp8/dsv4flash/1k1k/2p1d-ep16-mtp.yaml
+
+dsv4pro-fp8-mi355x-ep16-sglang:
+  model: sgl-project/DeepSeek-V4-Pro-FP8
+  model-prefix: dsv4pro
+  model_path: /it-share/model_coverage/models--sgl-project--DeepSeek-V4-Pro-FP8
+  runner: mi355x
+  precision: fp8
+  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-fp8/dsv4pro/1k1k/2p1d-ep16.yaml
+
+dsv4pro-fp8-mi355x-ep16-mtp-sglang:
+  model: sgl-project/DeepSeek-V4-Pro-FP8
+  model-prefix: dsv4pro
+  model_path: /it-share/model_coverage/models--sgl-project--DeepSeek-V4-Pro-FP8
+  runner: mi355x
+  precision: fp8
+  framework: sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+    - isl: 1024
+      osl: 1024
+      search-space:
+        # conc256 excluded: disagg-decode SWA hybrid pool retract->get_cpu_copy
+        # is an upstream NotImplementedError (crashes decode). See recipe.
+        - conc-list: [1, 8, 16, 32, 64, 128]
+          config_file: scripts/ci/slurm/recipes/mi355x-fp8/dsv4pro/1k1k/2p1d-ep16-mtp.yaml
+
+dsv4flash-fp4-mi355x-ep16-sglang:
+  model: deepseek-ai/DeepSeek-V4-Flash
+  model-prefix: dsv4flash
+  model_path: /it-share/model_coverage/models--deepseek-ai--DeepSeek-V4-Flash
+  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/dsv4flash/1k1k/2p1d-ep16.yaml
+
+dsv4flash-fp4-mi355x-ep16-mtp-sglang:
+  model: deepseek-ai/DeepSeek-V4-Flash
+  model-prefix: dsv4flash
+  model_path: /it-share/model_coverage/models--deepseek-ai--DeepSeek-V4-Flash
+  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/dsv4flash/1k1k/2p1d-ep16-mtp.yaml
+
+dsv4pro-fp4-mi355x-ep16-sglang:
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4pro
+  model_path: /it-share/model_coverage/models--deepseek-ai--DeepSeek-V4-Pro
+  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/dsv4pro/1k1k/2p1d-ep16.yaml
+
+dsv4pro-fp4-mi355x-ep16-mtp-sglang:
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4pro
+  model_path: /it-share/model_coverage/models--deepseek-ai--DeepSeek-V4-Pro
+  runner: mi355x
+  precision: fp4
+  framework: sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+    - isl: 1024
+      osl: 1024
+      search-space:
+        # conc256 excluded: disagg-decode SWA hybrid pool retract->get_cpu_copy
+        # is an upstream NotImplementedError (crashes decode). See recipe.
+        - conc-list: [1, 8, 16, 32, 64, 128]
+          config_file: scripts/ci/slurm/recipes/mi355x-fp4/dsv4pro/1k1k/2p1d-ep16-mtp.yaml
+
 # Kimi-K2.6 MXFP4 wide-EP16 2P1D: aiter MoE path, needs only the wide-EP launcher, not #32048.
 kimik26-mxfp4-mi355x-ep16-sglang:
   model: amd/Kimi-K2.6-MXFP4

scripts/ci/slurm/recipes/mi355x-fp4/dsv4flash/1k1k/2p1d-ep16-mtp.yaml

@@ -0,0 +1,107 @@
+# MI355X DeepSeek-V4-Flash FP4 4-node 2P1D disaggregation recipe — narrow-prefill EP8 + MTP
+# + wide-decode EP16 (Oren's config: wide EP only helps decode).
+#
+# Two prefill engines (EP8, one node each; the router fans requests across both) +
+# one decode engine (EP16) spanning 2 nodes. Still one logical P/D pair per role
+# group, 4 nodes total. nodes-per-engine = ceil(TP/8): prefill 8->1, decode 16->2,
+# so the launcher emits cross-node --nnodes/--node-rank/--dist-init-addr for the
+# decode engine only. Prefill EP8 keeps MoE all-to-all INTRA-node (XGMI, known
+# good); decode gets wide EP16 across nodes. Mismatched-TP KV (prefill TP8 ->
+# decode TP16) is layout-compatible for DeepSeek MLA (KV is a replicated per-token
+# latent, not head-sharded), and is carried over mori.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+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
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.15.post1-rocm720-mi35x-20260715
+  attention_backend: dsv4
+  # RoCE HCAs (8/node) for mori MoE all-to-all AND the P->D KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3,rdma4,rdma5,rdma6,rdma7
+  # Wide-EP MoE all-to-all backend (cross-node expert dispatch/combine).
+  moe_a2a_backend: mori
+  # KV P->D transfer backend (mori for both a2a and KV on this cluster).
+  kv_transfer_backend: mori
+  # Cross-node torch-distributed NIC for the wide decode engine's dist init.
+  dist_socket_ifname: eno0
+  # rocm720 0715 image needs the ROCm-7.0.0-alpha path OFF (validated).
+  rocm700a: 0
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  # Base defaults; the wide_ep block overrides mem-fraction / max-req per role.
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 131072
+  swa_full_tokens_ratio: 0.1
+  # Per-role wide-EP tuning (validated 2P1D run, job 13196). Prefill and decode
+  # differ, so these split what the single-node EP<=8 recipes leave symmetric.
+  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"
+    prefill_extra_flags: "--context-length 9217 --max-total-tokens 262144"
+    # Cap the decode KV pool: the ionic NIC rejects any single RDMA memory
+    # region over 4 GiB and mori registers each KV buffer as one region
+    # (no chunking).  Largest buffer = max_total_num_tokens * 256 B, so the
+    # ceiling is 16,777,216 tokens; MTP runs at 7,000,000 (the validated value,
+    # well under the ceiling and below every MTP leg's natural pool).
+    decode_extra_flags: "--max-total-tokens 7000000 --cuda-graph-bs 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 --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
+    # MTP decode capture drives the MoE dispatch to cuda_graph_bs * draft_tokens
+    # tokens/rank, so the decode buffers must clear that; 512/2048/1024 validated
+    # on mi355x (undersized values abort during decode cuda-graph capture).
+    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 (NextN head from the base model). Applied to
+# both prefill and decode.
+mtp:
+  enabled: true
+  num_steps: 3
+  eagle_topk: 1
+  num_draft_tokens: 4
+
+bench:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91

scripts/ci/slurm/recipes/mi355x-fp4/dsv4flash/1k1k/2p1d-ep16.yaml

@@ -0,0 +1,95 @@
+# MI355X DeepSeek-V4-Flash FP4 4-node 2P1D disaggregation recipe — narrow-prefill EP8
+# + wide-decode EP16 (Oren's config: wide EP only helps decode).
+#
+# Two prefill engines (EP8, one node each; the router fans requests across both) +
+# one decode engine (EP16) spanning 2 nodes. Still one logical P/D pair per role
+# group, 4 nodes total. nodes-per-engine = ceil(TP/8): prefill 8->1, decode 16->2,
+# so the launcher emits cross-node --nnodes/--node-rank/--dist-init-addr for the
+# decode engine only. Prefill EP8 keeps MoE all-to-all INTRA-node (XGMI, known
+# good); decode gets wide EP16 across nodes. Mismatched-TP KV (prefill TP8 ->
+# decode TP16) is layout-compatible for DeepSeek MLA (KV is a replicated per-token
+# latent, not head-sharded), and is carried over mori.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+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
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.15.post1-rocm720-mi35x-20260715
+  attention_backend: dsv4
+  # RoCE HCAs (8/node) for mori MoE all-to-all AND the P->D KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3,rdma4,rdma5,rdma6,rdma7
+  # Wide-EP MoE all-to-all backend (cross-node expert dispatch/combine).
+  moe_a2a_backend: mori
+  # KV P->D transfer backend (mori for both a2a and KV on this cluster).
+  kv_transfer_backend: mori
+  # Cross-node torch-distributed NIC for the wide decode engine's dist init.
+  dist_socket_ifname: eno0
+  # rocm720 0715 image needs the ROCm-7.0.0-alpha path OFF (validated).
+  rocm700a: 0
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  # Base defaults; the wide_ep block overrides mem-fraction / max-req per role.
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 131072
+  swa_full_tokens_ratio: 0.1
+  # Per-role wide-EP tuning (validated 2P1D run, job 13196). Prefill and decode
+  # differ, so these split what the single-node EP<=8 recipes leave symmetric.
+  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"
+    prefill_extra_flags: "--context-length 9217 --max-total-tokens 262144"
+    # Cap the decode KV pool: the ionic NIC rejects any single RDMA memory
+    # region over 4 GiB and mori registers each KV buffer as one region
+    # (no chunking).  Largest buffer = max_total_num_tokens * 256 B, so the
+    # ceiling is 16,777,216 tokens; 16,000,000 leaves headroom.
+    decode_extra_flags: "--max-total-tokens 16000000 --cuda-graph-bs 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 --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:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91

scripts/ci/slurm/recipes/mi355x-fp4/dsv4pro/1k1k/2p1d-ep16-mtp.yaml

@@ -0,0 +1,107 @@
+# MI355X DeepSeek-V4-Pro FP4 4-node 2P1D disaggregation recipe — narrow-prefill EP8 + MTP
+# + wide-decode EP16 (Oren's config: wide EP only helps decode).
+#
+# Two prefill engines (EP8, one node each; the router fans requests across both) +
+# one decode engine (EP16) spanning 2 nodes. Still one logical P/D pair per role
+# group, 4 nodes total. nodes-per-engine = ceil(TP/8): prefill 8->1, decode 16->2,
+# so the launcher emits cross-node --nnodes/--node-rank/--dist-init-addr for the
+# decode engine only. Prefill EP8 keeps MoE all-to-all INTRA-node (XGMI, known
+# good); decode gets wide EP16 across nodes. Mismatched-TP KV (prefill TP8 ->
+# decode TP16) is layout-compatible for DeepSeek MLA (KV is a replicated per-token
+# latent, not head-sharded), and is carried over mori.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+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
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.15.post1-rocm720-mi35x-20260715
+  attention_backend: dsv4
+  # RoCE HCAs (8/node) for mori MoE all-to-all AND the P->D KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3,rdma4,rdma5,rdma6,rdma7
+  # Wide-EP MoE all-to-all backend (cross-node expert dispatch/combine).
+  moe_a2a_backend: mori
+  # KV P->D transfer backend (mori for both a2a and KV on this cluster).
+  kv_transfer_backend: mori
+  # Cross-node torch-distributed NIC for the wide decode engine's dist init.
+  dist_socket_ifname: eno0
+  # rocm720 0715 image needs the ROCm-7.0.0-alpha path OFF (validated).
+  rocm700a: 0
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  # Base defaults; the wide_ep block overrides mem-fraction / max-req per role.
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 131072
+  swa_full_tokens_ratio: 0.1
+  # Per-role wide-EP tuning (validated 2P1D run, job 13196). Prefill and decode
+  # differ, so these split what the single-node EP<=8 recipes leave symmetric.
+  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"
+    prefill_extra_flags: "--context-length 9217 --max-total-tokens 262144"
+    # Cap the decode KV pool: the ionic NIC rejects any single RDMA memory
+    # region over 4 GiB and mori registers each KV buffer as one region
+    # (no chunking).  Largest buffer = max_total_num_tokens * 256 B, so the
+    # ceiling is 16,777,216 tokens; MTP runs at 7,000,000 (the validated value,
+    # well under the ceiling and below every MTP leg's natural pool).
+    decode_extra_flags: "--max-total-tokens 7000000 --cuda-graph-bs 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 --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
+    # MTP decode capture drives the MoE dispatch to cuda_graph_bs * draft_tokens
+    # tokens/rank, so the decode buffers must clear that; 512/2048/1024 validated
+    # on mi355x (undersized values abort during decode cuda-graph capture).
+    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 (NextN head from the base model). Applied to
+# both prefill and decode.
+mtp:
+  enabled: true
+  num_steps: 3
+  eagle_topk: 1
+  num_draft_tokens: 4
+
+bench:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91

scripts/ci/slurm/recipes/mi355x-fp4/dsv4pro/1k1k/2p1d-ep16.yaml

@@ -0,0 +1,95 @@
+# MI355X DeepSeek-V4-Pro FP4 4-node 2P1D disaggregation recipe — narrow-prefill EP8
+# + wide-decode EP16 (Oren's config: wide EP only helps decode).
+#
+# Two prefill engines (EP8, one node each; the router fans requests across both) +
+# one decode engine (EP16) spanning 2 nodes. Still one logical P/D pair per role
+# group, 4 nodes total. nodes-per-engine = ceil(TP/8): prefill 8->1, decode 16->2,
+# so the launcher emits cross-node --nnodes/--node-rank/--dist-init-addr for the
+# decode engine only. Prefill EP8 keeps MoE all-to-all INTRA-node (XGMI, known
+# good); decode gets wide EP16 across nodes. Mismatched-TP KV (prefill TP8 ->
+# decode TP16) is layout-compatible for DeepSeek MLA (KV is a replicated per-token
+# latent, not head-sharded), and is carried over mori.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+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
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.15.post1-rocm720-mi35x-20260715
+  attention_backend: dsv4
+  # RoCE HCAs (8/node) for mori MoE all-to-all AND the P->D KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3,rdma4,rdma5,rdma6,rdma7
+  # Wide-EP MoE all-to-all backend (cross-node expert dispatch/combine).
+  moe_a2a_backend: mori
+  # KV P->D transfer backend (mori for both a2a and KV on this cluster).
+  kv_transfer_backend: mori
+  # Cross-node torch-distributed NIC for the wide decode engine's dist init.
+  dist_socket_ifname: eno0
+  # rocm720 0715 image needs the ROCm-7.0.0-alpha path OFF (validated).
+  rocm700a: 0
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  # Base defaults; the wide_ep block overrides mem-fraction / max-req per role.
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 131072
+  swa_full_tokens_ratio: 0.1
+  # Per-role wide-EP tuning (validated 2P1D run, job 13196). Prefill and decode
+  # differ, so these split what the single-node EP<=8 recipes leave symmetric.
+  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"
+    prefill_extra_flags: "--context-length 9217 --max-total-tokens 262144"
+    # Cap the decode KV pool: the ionic NIC rejects any single RDMA memory
+    # region over 4 GiB and mori registers each KV buffer as one region
+    # (no chunking).  Largest buffer = max_total_num_tokens * 256 B, so the
+    # ceiling is 16,777,216 tokens; 16,000,000 leaves headroom.
+    decode_extra_flags: "--max-total-tokens 16000000 --cuda-graph-bs 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 --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:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91

scripts/ci/slurm/recipes/mi355x-fp8/dsv4flash/1k1k/2p1d-ep16-mtp.yaml

@@ -0,0 +1,107 @@
+# MI355X DeepSeek-V4-Flash FP8 4-node 2P1D disaggregation recipe — narrow-prefill EP8 + MTP
+# + wide-decode EP16 (Oren's config: wide EP only helps decode).
+#
+# Two prefill engines (EP8, one node each; the router fans requests across both) +
+# one decode engine (EP16) spanning 2 nodes. Still one logical P/D pair per role
+# group, 4 nodes total. nodes-per-engine = ceil(TP/8): prefill 8->1, decode 16->2,
+# so the launcher emits cross-node --nnodes/--node-rank/--dist-init-addr for the
+# decode engine only. Prefill EP8 keeps MoE all-to-all INTRA-node (XGMI, known
+# good); decode gets wide EP16 across nodes. Mismatched-TP KV (prefill TP8 ->
+# decode TP16) is layout-compatible for DeepSeek MLA (KV is a replicated per-token
+# latent, not head-sharded), and is carried over mori.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+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
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.15.post1-rocm720-mi35x-20260715
+  attention_backend: dsv4
+  # RoCE HCAs (8/node) for mori MoE all-to-all AND the P->D KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3,rdma4,rdma5,rdma6,rdma7
+  # Wide-EP MoE all-to-all backend (cross-node expert dispatch/combine).
+  moe_a2a_backend: mori
+  # KV P->D transfer backend (mori for both a2a and KV on this cluster).
+  kv_transfer_backend: mori
+  # Cross-node torch-distributed NIC for the wide decode engine's dist init.
+  dist_socket_ifname: eno0
+  # rocm720 0715 image needs the ROCm-7.0.0-alpha path OFF (validated).
+  rocm700a: 0
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  # Base defaults; the wide_ep block overrides mem-fraction / max-req per role.
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 131072
+  swa_full_tokens_ratio: 0.1
+  # Per-role wide-EP tuning (validated 2P1D run, job 13196). Prefill and decode
+  # differ, so these split what the single-node EP<=8 recipes leave symmetric.
+  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"
+    prefill_extra_flags: "--context-length 9217 --max-total-tokens 262144"
+    # Cap the decode KV pool: the ionic NIC rejects any single RDMA memory
+    # region over 4 GiB and mori registers each KV buffer as one region
+    # (no chunking).  Largest buffer = max_total_num_tokens * 256 B, so the
+    # ceiling is 16,777,216 tokens; MTP runs at 7,000,000 (the validated value,
+    # well under the ceiling and below every MTP leg's natural pool).
+    decode_extra_flags: "--max-total-tokens 7000000 --cuda-graph-bs 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 --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
+    # MTP decode capture drives the MoE dispatch to cuda_graph_bs * draft_tokens
+    # tokens/rank, so the decode buffers must clear that; 512/2048/1024 validated
+    # on mi355x (undersized values abort during decode cuda-graph capture).
+    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 (NextN head from the base model). Applied to
+# both prefill and decode.
+mtp:
+  enabled: true
+  num_steps: 3
+  eagle_topk: 1
+  num_draft_tokens: 4
+
+bench:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91

scripts/ci/slurm/recipes/mi355x-fp8/dsv4flash/1k1k/2p1d-ep16.yaml

@@ -0,0 +1,95 @@
+# MI355X DeepSeek-V4-Flash FP8 4-node 2P1D disaggregation recipe — narrow-prefill EP8
+# + wide-decode EP16 (Oren's config: wide EP only helps decode).
+#
+# Two prefill engines (EP8, one node each; the router fans requests across both) +
+# one decode engine (EP16) spanning 2 nodes. Still one logical P/D pair per role
+# group, 4 nodes total. nodes-per-engine = ceil(TP/8): prefill 8->1, decode 16->2,
+# so the launcher emits cross-node --nnodes/--node-rank/--dist-init-addr for the
+# decode engine only. Prefill EP8 keeps MoE all-to-all INTRA-node (XGMI, known
+# good); decode gets wide EP16 across nodes. Mismatched-TP KV (prefill TP8 ->
+# decode TP16) is layout-compatible for DeepSeek MLA (KV is a replicated per-token
+# latent, not head-sharded), and is carried over mori.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+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
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.15.post1-rocm720-mi35x-20260715
+  attention_backend: dsv4
+  # RoCE HCAs (8/node) for mori MoE all-to-all AND the P->D KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3,rdma4,rdma5,rdma6,rdma7
+  # Wide-EP MoE all-to-all backend (cross-node expert dispatch/combine).
+  moe_a2a_backend: mori
+  # KV P->D transfer backend (mori for both a2a and KV on this cluster).
+  kv_transfer_backend: mori
+  # Cross-node torch-distributed NIC for the wide decode engine's dist init.
+  dist_socket_ifname: eno0
+  # rocm720 0715 image needs the ROCm-7.0.0-alpha path OFF (validated).
+  rocm700a: 0
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  # Base defaults; the wide_ep block overrides mem-fraction / max-req per role.
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 131072
+  swa_full_tokens_ratio: 0.1
+  # Per-role wide-EP tuning (validated 2P1D run, job 13196). Prefill and decode
+  # differ, so these split what the single-node EP<=8 recipes leave symmetric.
+  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"
+    prefill_extra_flags: "--context-length 9217 --max-total-tokens 262144"
+    # Cap the decode KV pool: the ionic NIC rejects any single RDMA memory
+    # region over 4 GiB and mori registers each KV buffer as one region
+    # (no chunking).  Largest buffer = max_total_num_tokens * 256 B, so the
+    # ceiling is 16,777,216 tokens; 16,000,000 leaves headroom.
+    decode_extra_flags: "--max-total-tokens 16000000 --cuda-graph-bs 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 --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:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91

scripts/ci/slurm/recipes/mi355x-fp8/dsv4pro/1k1k/2p1d-ep16-mtp.yaml

@@ -0,0 +1,107 @@
+# MI355X DeepSeek-V4-Pro FP8 4-node 2P1D disaggregation recipe — narrow-prefill EP8 + MTP
+# + wide-decode EP16 (Oren's config: wide EP only helps decode).
+#
+# Two prefill engines (EP8, one node each; the router fans requests across both) +
+# one decode engine (EP16) spanning 2 nodes. Still one logical P/D pair per role
+# group, 4 nodes total. nodes-per-engine = ceil(TP/8): prefill 8->1, decode 16->2,
+# so the launcher emits cross-node --nnodes/--node-rank/--dist-init-addr for the
+# decode engine only. Prefill EP8 keeps MoE all-to-all INTRA-node (XGMI, known
+# good); decode gets wide EP16 across nodes. Mismatched-TP KV (prefill TP8 ->
+# decode TP16) is layout-compatible for DeepSeek MLA (KV is a replicated per-token
+# latent, not head-sharded), and is carried over mori.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+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
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.15.post1-rocm720-mi35x-20260715
+  attention_backend: dsv4
+  # RoCE HCAs (8/node) for mori MoE all-to-all AND the P->D KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3,rdma4,rdma5,rdma6,rdma7
+  # Wide-EP MoE all-to-all backend (cross-node expert dispatch/combine).
+  moe_a2a_backend: mori
+  # KV P->D transfer backend (mori for both a2a and KV on this cluster).
+  kv_transfer_backend: mori
+  # Cross-node torch-distributed NIC for the wide decode engine's dist init.
+  dist_socket_ifname: eno0
+  # rocm720 0715 image needs the ROCm-7.0.0-alpha path OFF (validated).
+  rocm700a: 0
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  # Base defaults; the wide_ep block overrides mem-fraction / max-req per role.
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 131072
+  swa_full_tokens_ratio: 0.1
+  # Per-role wide-EP tuning (validated 2P1D run, job 13196). Prefill and decode
+  # differ, so these split what the single-node EP<=8 recipes leave symmetric.
+  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"
+    prefill_extra_flags: "--context-length 9217 --max-total-tokens 262144"
+    # Cap the decode KV pool: the ionic NIC rejects any single RDMA memory
+    # region over 4 GiB and mori registers each KV buffer as one region
+    # (no chunking).  Largest buffer = max_total_num_tokens * 256 B, so the
+    # ceiling is 16,777,216 tokens; MTP runs at 7,000,000 (the validated value,
+    # well under the ceiling and below every MTP leg's natural pool).
+    decode_extra_flags: "--max-total-tokens 7000000 --cuda-graph-bs 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 --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
+    # MTP decode capture drives the MoE dispatch to cuda_graph_bs * draft_tokens
+    # tokens/rank, so the decode buffers must clear that; 512/2048/1024 validated
+    # on mi355x (undersized values abort during decode cuda-graph capture).
+    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 (NextN head from the base model). Applied to
+# both prefill and decode.
+mtp:
+  enabled: true
+  num_steps: 3
+  eagle_topk: 1
+  num_draft_tokens: 4
+
+bench:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91

scripts/ci/slurm/recipes/mi355x-fp8/dsv4pro/1k1k/2p1d-ep16.yaml

@@ -0,0 +1,95 @@
+# MI355X DeepSeek-V4-Pro FP8 4-node 2P1D disaggregation recipe — narrow-prefill EP8
+# + wide-decode EP16 (Oren's config: wide EP only helps decode).
+#
+# Two prefill engines (EP8, one node each; the router fans requests across both) +
+# one decode engine (EP16) spanning 2 nodes. Still one logical P/D pair per role
+# group, 4 nodes total. nodes-per-engine = ceil(TP/8): prefill 8->1, decode 16->2,
+# so the launcher emits cross-node --nnodes/--node-rank/--dist-init-addr for the
+# decode engine only. Prefill EP8 keeps MoE all-to-all INTRA-node (XGMI, known
+# good); decode gets wide EP16 across nodes. Mismatched-TP KV (prefill TP8 ->
+# decode TP16) is layout-compatible for DeepSeek MLA (KV is a replicated per-token
+# latent, not head-sharded), and is carried over mori.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+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
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.15.post1-rocm720-mi35x-20260715
+  attention_backend: dsv4
+  # RoCE HCAs (8/node) for mori MoE all-to-all AND the P->D KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3,rdma4,rdma5,rdma6,rdma7
+  # Wide-EP MoE all-to-all backend (cross-node expert dispatch/combine).
+  moe_a2a_backend: mori
+  # KV P->D transfer backend (mori for both a2a and KV on this cluster).
+  kv_transfer_backend: mori
+  # Cross-node torch-distributed NIC for the wide decode engine's dist init.
+  dist_socket_ifname: eno0
+  # rocm720 0715 image needs the ROCm-7.0.0-alpha path OFF (validated).
+  rocm700a: 0
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  # Base defaults; the wide_ep block overrides mem-fraction / max-req per role.
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 131072
+  swa_full_tokens_ratio: 0.1
+  # Per-role wide-EP tuning (validated 2P1D run, job 13196). Prefill and decode
+  # differ, so these split what the single-node EP<=8 recipes leave symmetric.
+  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"
+    prefill_extra_flags: "--context-length 9217 --max-total-tokens 262144"
+    # Cap the decode KV pool: the ionic NIC rejects any single RDMA memory
+    # region over 4 GiB and mori registers each KV buffer as one region
+    # (no chunking).  Largest buffer = max_total_num_tokens * 256 B, so the
+    # ceiling is 16,777,216 tokens; 16,000,000 leaves headroom.
+    decode_extra_flags: "--max-total-tokens 16000000 --cuda-graph-bs 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 --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:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91