[AMD] dsv4-fp4-mi355x-atom: latest atom-dev image + opt recipe / DSv4-Pro FP4 MI355X ATOM：升级最新 atom-dev 镜像 + 优化配方

benchmarks/single_node/fixed_seq_len/dsv4_fp4_mi355x_atom.sh

@@ -27,11 +27,8 @@ if [ "$DP_ATTENTION" = "true" ]; then
     if [ "$EP_SIZE" -gt 1 ]; then #DP+EP
         PARALLEL_ARGS=(-tp "$TP" --enable-expert-parallel --enable-dp-attention )
     else #DPA+TP
-        #DPA+TP+TBO
-        if [ "$ISL" -eq 1024 ] && [ "$OSL" -eq 1024 ] && [ "$CONC" -ge 1024 ]; then
-            PARALLEL_ARGS=(-tp "$TP" --enable-dp-attention --enable-tbo)
-            export GPU_MAX_HW_QUEUES=5
-        elif [ "$ISL" -eq 8192 ] && [ "$OSL" -eq 1024 ] && [ "$CONC" -ge 256 ]; then
+        #DPA+TP+TBO (opt: TBO on for dp-attn cells at conc>=64, no per-scenario gate)
+        if [ "$CONC" -ge 64 ]; then
             PARALLEL_ARGS=(-tp "$TP" --enable-dp-attention --enable-tbo)
             export GPU_MAX_HW_QUEUES=5
         else
@@ -40,6 +37,14 @@ if [ "$DP_ATTENTION" = "true" ]; then
     fi
 fi 
 
+# max_req=conc for every dp-on cell (mandatory: dp-attention keeps a full KV pool
+# per rank, so the large default max_num_seqs OOMs even at low conc like c16/c32)
+# and for mid/high conc (conc>=64). dp-off low conc uses the ATOM default
+# (dev: default is on-par or ~4% better at very low conc, e.g. c2).
+if [ "$DP_ATTENTION" = "true" ] || [ "$CONC" -ge 64 ]; then
+    PARALLEL_ARGS+=(--max-num-seqs "$CONC")
+fi
+
 BENCHMARK_MAX_MODEL_LEN="$MAX_MODEL_LEN"
 
 if [ "${EVAL_ONLY}" = "true" ]; then

configs/amd-master.yaml

@@ -1405,7 +1405,7 @@ dsv4-fp4-mi355x-vllm-mtp:
       - { tp: 8, conc-start: 4, conc-end: 512, spec-decoding: mtp }
 
 dsv4-fp4-mi355x-atom:
-  image: rocm/atom-dev:nightly_202606161823
+  image: rocm/atom-dev:nightly_202607231538
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: mi355x
@@ -1417,12 +1417,14 @@ dsv4-fp4-mi355x-atom:
     - isl: 8192
       osl: 1024
       search-space:
-        # conc4-64, TP8
-        # conc128, DPA
-        # conc256-2048, DPA TBO
-      - { tp: 4, ep: 1, conc-list: [8, 16, 32, 64] }
-      - { tp: 8, ep: 1, conc-list: [1, 2, 4, 8, 16, 32, 64] }
-      - { tp: 8, ep: 1, dp-attn: true, conc-start: 128, conc-end: 2048 }
+      # conc<64: tp4 dp-off (best at low conc); conc>=64: dp-on + TBO.
+      # tp8 dp-off low-conc kept as redundancy to recheck the crossover on the new image.
+      # dp-on cells always set max_req=conc (recipe .sh) to avoid OOM from the large
+      # default max_num_seqs (dp-attn keeps a full KV pool per rank); TBO stays conc>=64.
+      - { tp: 4, ep: 1, dp-attn: false, conc-start: 1,  conc-end: 64 }
+      - { tp: 4, ep: 1, dp-attn: true,  conc-start: 16, conc-end: 128 }
+      - { tp: 8, ep: 1, dp-attn: false, conc-start: 1,  conc-end: 64 }
+      - { tp: 8, ep: 1, dp-attn: true,  conc-start: 64, conc-end: 2048 }
 
 dsv4-fp4-mi355x-atom-mtp:
   image: rocm/atom:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.3

perf-changelog.yaml

@@ -5086,6 +5086,12 @@
     - "Bump SGLang container image from lmsysorg/sglang:v0.5.12-cu130 to lmsysorg/sglang:v0.5.15.post1-cu130 (https://github.com/sgl-project/sglang/releases/tag/v0.5.15.post1)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2313
 
+- config-keys:
+    - dsv4-fp4-mi355x-atom
+  description:
+    - "Re-sweep dsv4-fp4-mi355x-atom on latest atom-dev nightly (nightly_202607231538, was 202606161823). Recipe: max_num_seqs=conc for all dp-on cells (avoids dp-attn OOM) and conc>=64; TBO on dp-attn cells at conc>=64 (was per-scenario gate 256/1024). 8k1k grid adds tp8 dp-off low-conc redundancy to recheck the crossover on the new image."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2327
+
 - config-keys:
     - minimaxm3-fp8-b200-vllm
     - minimaxm3-fp8-b300-vllm