[AMD][DI][CI] 2/N Add DSV4 DP8/EP8 and MTP MI355X 1P1D nightly recipes (#29784)

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

scripts/ci/slurm/recipes/mi355x-fp8/dsv4flash/1k1k/1p1d-dp8ep8-mtp.yaml

@@ -0,0 +1,61 @@
+# MI355X DeepSeek-V4-Flash-FP8 2-node 1P1D disaggregation recipe — DP8 + narrow EP8 + MTP.
+#
+# 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: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  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: 8192
+  swa_full_tokens_ratio: 0.1
+
+# MTP / EAGLE speculative decoding (NextN head from the base model). Applied to
+# both prefill and decode. Flags mirror
+# test/registered/amd/test_deepseek_v4_pro_fp4_mtp.py.
+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/1p1d-dp8ep8.yaml

@@ -0,0 +1,52 @@
+# MI355X DeepSeek-V4-Flash-FP8 2-node 1P1D disaggregation recipe — DP8 + narrow EP8.
+#
+# 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: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  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: 8192
+  swa_full_tokens_ratio: 0.1
+
+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/1p1d-mtp.yaml

@@ -0,0 +1,61 @@
+# MI355X DeepSeek-V4-Flash-FP8 2-node 1P1D disaggregation recipe — TP8 + MTP.
+#
+# 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: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 1
+      data-parallel-size: 1
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 1
+      data-parallel-size: 1
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  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: 8192
+  swa_full_tokens_ratio: 0.1
+
+# MTP / EAGLE speculative decoding (NextN head from the base model). Applied to
+# both prefill and decode. Flags mirror
+# test/registered/amd/test_deepseek_v4_pro_fp4_mtp.py.
+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/1p1d-dp8ep8-mtp.yaml

@@ -0,0 +1,61 @@
+# MI355X DeepSeek-V4-Pro-FP8 2-node 1P1D disaggregation recipe — DP8 + narrow EP8 + MTP.
+#
+# 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: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  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: 8192
+  swa_full_tokens_ratio: 0.1
+
+# MTP / EAGLE speculative decoding (NextN head from the base model). Applied to
+# both prefill and decode. Flags mirror
+# test/registered/amd/test_deepseek_v4_pro_fp4_mtp.py.
+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/1p1d-dp8ep8.yaml

@@ -0,0 +1,52 @@
+# MI355X DeepSeek-V4-Pro-FP8 2-node 1P1D disaggregation recipe — DP8 + narrow EP8.
+#
+# 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: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  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: 8192
+  swa_full_tokens_ratio: 0.1
+
+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/1p1d-mtp.yaml

@@ -0,0 +1,68 @@
+# MI355X DeepSeek-V4-Pro-FP8 2-node 1P1D disaggregation recipe — TP8 + MTP.
+#
+# 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: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 1
+      data-parallel-size: 1
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 1
+      data-parallel-size: 1
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  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: 8192
+  swa_full_tokens_ratio: 0.1
+
+# MTP / EAGLE speculative decoding (NextN head from the base model). Applied to
+# both prefill and decode. Flags mirror
+# test/registered/amd/test_deepseek_v4_pro_fp4_mtp.py.
+mtp:
+  enabled: true
+  num_steps: 3
+  eagle_topk: 1
+  num_draft_tokens: 4
+
+bench:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  # conc256 is excluded for this leg only: at conc256 the disagg-decode SWA
+  # hybrid KV pool fills, the scheduler retracts running requests, and the
+  # retract->offload_kv_cache path calls get_cpu_copy() which is unimplemented
+  # for the SWA hybrid pool (raises NotImplementedError, crashes the decode
+  # scheduler). Raising swa_full_tokens_ratio (tried up to 0.3) does not help
+  # -- usage climbs to fill the larger budget and still retracts. Until the
+  # upstream get_cpu_copy stub is implemented, cap this leg at 128.
+  concurrencies: [1, 8, 16, 32, 64, 128]
+  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