[WIP] Chore/agentx v0.3

.github/configs/amd-master.yaml

@@ -323,6 +323,21 @@ qwen3.5-fp8-mi355x-sglang-agentic:
       search-space:
       - { tp: 8, ep: 1, offloading: none, conc-list: [1, 2, 4, 8, 16, 32] }
 
+qwen3.5-fp8-mi355x-sglang-agentic-hicache:
+  image: lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260521
+  model: Qwen/Qwen3.5-397B-A17B-FP8
+  model-prefix: qwen3.5
+  runner: mi355x
+  precision: fp8
+  framework: sglang
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - duration: 1800
+      search-space:
+      - { tp: 8, ep: 1, offloading: none, conc-list: [1, 2, 4, 8, 16, 32] }
+      - { tp: 8, ep: 1, offloading: hicache, conc-list: [16, 32, 48, 64] }
+
 qwen3.5-fp8-mi355x-atom:
   image: rocm/atom:rocm7.2.3_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom20260511
   model: Qwen/Qwen3.5-397B-A17B-FP8
@@ -653,10 +668,6 @@ kimik2.5-fp4-mi355x-vllm:
 # its fixed-seq-len sweep is unaffected.
 #   - image: 'vllm/vllm-openai-rocm:v0.18.0' -> 'vllm/vllm-openai-rocm:v0.21.0'
 kimik2.5-fp4-mi355x-vllm-agentic:
-  # v0.21.0 (released 2026-05-14) supersedes the prior nightly pin
-  # (51f22dcf...) which was carrying the SimpleCPUOffloadConnector ROCm
-  # cpu_offload_blocks > 0 fix. v0.21.0 is much newer than that fix and
-  # includes all subsequent ROCm offload work.
   image: vllm/vllm-openai-rocm:v0.21.0
   model: amd/Kimi-K2.5-MXFP4
   model-prefix: kimik2.5
@@ -669,16 +680,9 @@ kimik2.5-fp4-mi355x-vllm-agentic:
     - duration: 1800
       search-space:
       - { tp: 8, offloading: none, conc-list: [1, 2, 4, 8, 16, 24, 32, 40, 48] }
-      # CPU offload only above the KV cliff. Lower concurrencies fit
-      # entirely on-GPU, so paying the offload-path overhead there would
-      # just slow them down without measuring anything new.
-      - { tp: 8, offloading: cpu,  conc-list: [32, 40, 48, 56] }
-      # TP=4 probe: half-node layout doubles per-GPU weight footprint
-      # (~62 GB on MI355X's 288 GB HBM, plenty of headroom). Restrict to
-      # cliff-region concurrencies on both offload modes so we can directly
-      # compare TP=4 vs TP=8 at the same conc points.
+      - { tp: 8, offloading: lmcache,  conc-list: [32, 40, 48, 56] }
       - { tp: 4, offloading: none, conc-list: [16, 24, 32, 40] }
-      - { tp: 4, offloading: cpu,  conc-list: [16, 24, 32, 40] }
+      - { tp: 4, offloading: lmcache,  conc-list: [16, 24, 32, 40] }
 
 kimik2.5-fp4-mi355x-atom:
   image: rocm/atom:rocm7.2.3_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom20260511
@@ -701,6 +705,22 @@ kimik2.5-fp4-mi355x-atom:
       - { tp: 8, conc-start: 4, conc-end: 128 }
       - { tp: 4, conc-start: 4, conc-end: 128 }
 
+dsv4-fp4-mi355x-vllm-agentic:
+  image: vllm/vllm-openai-rocm:v0.21.0
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: mi355x
+  precision: fp4
+  framework: vllm
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - duration: 1800
+      search-space:
+      - { tp: 8, offloading: none, conc-list: [1, 2, 4] }
+      - { tp: 4, offloading: none, conc-list: [1, 2, 4, 8, 10, 12, 16] }
+      - { tp: 4, ep: 4, dp-attn: true, offloading: none, conc-list: [16, 24, 32, 40, 48] }
+
 minimaxm2.5-fp8-mi355x-vllm:
   image: vllm/vllm-openai-rocm:v0.21.0
   model: MiniMaxAI/MiniMax-M2.5
@@ -1833,6 +1853,29 @@ dsv4-fp4-mi355x-sglang:
       - { tp: 8, dp-attn: true, conc-start: 64, conc-end: 2048 }
       - { tp: 8, dp-attn: false, conc-start: 1, conc-end: 32 }
 
+# Diverged from dsv4-fp4-mi355x-sglang (agentic-coding sibling). Reasons below;
+# the original dsv4-fp4-mi355x-sglang entry is left identical to origin/main so
+# its fixed-seq-len sweep is unaffected.
+#   - scenarios: replaced fixed-seq-len with agentic-coding.
+# Image is identical to the base entry (rocm/sgl-dev DSv4 build).
+# CONC ranges mirror dsv4-fp4-b200-vllm-agentic for cross-hardware
+# comparability. Offload sweep is none-only (SGLang has no equivalent of
+# vLLM's SimpleCPUOffloadConnector path that we exercise on b200).
+dsv4-fp4-mi355x-sglang-agentic:
+  image: rocm/sgl-dev:rocm720-mi35x-0363e6c-20260509-DSv4
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: mi355x
+  precision: fp4
+  framework: sglang
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - duration: 1800
+      search-space:
+      - { tp: 8, offloading: none, conc-list: [16, 32, 64] }
+      - { tp: 8, dp-attn: true, offloading: none, conc-list: [64, 128, 256] }
+
 # DSv4 on MI355X via vLLM, using the official vllm/vllm-openai-rocm
 # nightly image. DSv4 base ROCm support (vllm-project/vllm#40871) merged
 # on 2026-05-05, so any nightly built after that includes the