fix(dsv4): use all 16 dies with TP4/DP4 + TP8/DP2 + TP16/DP1

The A3 910C has 16 dies (8 cards x 2 dies/card), and vllm-ascend's
tensor-parallel-size / data-parallel-size address DIES, not cards. The
old configs (2/4, 4/2, 8/1) all used only 8 dies = 4 cards, leaving half
the node idle. Switch to configs that use all 16 dies, and add per-TP
parameter overrides since each TP has a very different per-die memory
budget.

Why the old configs were wrong:
- A3 TP=4 (4 dies) is the per-CARD equivalent of H20 TP=4 (4 cards),
  not a fair comparison. An A3 node has 2x the compute of an 8-card H20,
  so benchmarking at 8 dies understates the A3 and never exercises the
  cross-card / full-NVLink topology.
- TP*DP must equal 16 to use all dies on the A3 node.

New PARALLEL_CONFIGS (all use 16 dies):
- TP=4  DP=4: 4 dies/replica x 4 replicas, ~39 GiB weights/die
- TP=8  DP=2: 8 dies/replica x 2 replicas, ~35 GiB weights/die
- TP=16 DP=1: 16 dies/replica x 1 replica,  ~17.5 GiB weights/die

Per-TP overrides (mirrors the glm52 6c81183 pattern):
- TP4:  max_model_len=32768  max_num_seqs=128  (tight KV cache)
- TP8:  max_model_len=65536  max_num_seqs=256  (balanced)
- TP16: max_model_len=131072 max_num_seqs=256  (max KV cache)

Changes:
- config.env: PARALLEL_CONFIGS -> "4 4"/"8 2"/"16 1"; add TP4_/TP8_/TP16_
  env vars for per-TP gpu_mem_util/max_model_len/max_num_seqs.
- start_vllm_docker.sh: case "$TP" overrides the three params after
  sourcing config.env, so the actual launch args match the per-TP budget.
- run_adaptive_concurrency_add16.sh: engine_build_server_args gets the
  same case "$TP" so the recorded server_cmd.txt stays consistent with
  the real launch.

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/config.env

@@ -49,9 +49,12 @@ export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-0,1,2,3,4,5,6,7}"
 RUNTIME_BASE="${RUNTIME_BASE:-${SCRIPT_DIR}/runtime}"
 
 # Parallel configurations to test. Format: "TP DP"
-# DSV4-Flash w8a8 routed-expert weights ~280 GiB total. With expert-parallel the
+# A3 910C has 16 dies (8 cards x 2 dies/card). TP/DP address dies, so TP*DP=16
-# weight load is sharded across TP dies, but w8a8 still leaves a heavy per-die
+# uses all dies. DSV4-Flash w8a8 weights ~280 GiB total, sharded across TP dies:
-# footprint. TP*DP must equal 16 (8 cards x 2 dies).
+#   TP=4  DP=4 -> 4 dies/replica x 4 replicas (39 GiB/die, tight KV cache)
+#   TP=8  DP=2 -> 8 dies/replica x 2 replicas (35 GiB/die, balanced)
+#   TP=16 DP=1 -> 16 dies/replica x 1 replica (17.5 GiB/die, max KV cache)
+# Note: A3 TP=4 == H20 TP=4 per-card-equivalent (A3 has 2 dies/card).
 #   TP=2, DP=4 -> 4 dies/replica x 4 replicas (smallest TP, most replicas)
 #   TP=4, DP=2 -> 4 dies/replica x 2 replicas
 #   TP=8, DP=1 -> 8 dies/replica x 1 replica  (largest TP, max weight sharding)
@@ -62,9 +65,9 @@ if [[ -n "${PARALLEL_CONFIGS_STR:-}" ]]; then
   done
 else
   declare -a PARALLEL_CONFIGS=(
-    "2 4"
+    "4 4"
-    "4 2"
+    "8 2"
-    "8 1"
+    "16 1"
   )
 fi
 
@@ -81,6 +84,21 @@ BLOCK_SIZE="${BLOCK_SIZE:-128}"
 MAX_MODEL_LEN="${MAX_MODEL_LEN:-131072}"
 MAX_NUM_SEQS="${MAX_NUM_SEQS:-256}"
 
+# Per-TP parameter overrides (applied in start_vllm_docker.sh via case $TP).
+# DSV4-Flash w8a8 weight per die = ~280GiB / TP. Each die has 64GB HBM.
+#   TP=4:  ~39 GiB/die weights -> 25 GiB for KV cache; cap context to avoid OOM
+#   TP=8:  ~35 GiB/die weights -> 29 GiB for KV cache; balanced
+#   TP=16: ~17.5 GiB/die weights -> 46 GiB for KV cache; full context
+TP4_GPU_MEMORY_UTILIZATION="${TP4_GPU_MEMORY_UTILIZATION:-0.9}"
+TP4_MAX_MODEL_LEN="${TP4_MAX_MODEL_LEN:-32768}"
+TP4_MAX_NUM_SEQS="${TP4_MAX_NUM_SEQS:-128}"
+TP8_GPU_MEMORY_UTILIZATION="${TP8_GPU_MEMORY_UTILIZATION:-0.9}"
+TP8_MAX_MODEL_LEN="${TP8_MAX_MODEL_LEN:-65536}"
+TP8_MAX_NUM_SEQS="${TP8_MAX_NUM_SEQS:-256}"
+TP16_GPU_MEMORY_UTILIZATION="${TP16_GPU_MEMORY_UTILIZATION:-0.9}"
+TP16_MAX_MODEL_LEN="${TP16_MAX_MODEL_LEN:-131072}"
+TP16_MAX_NUM_SEQS="${TP16_MAX_NUM_SEQS:-256}"
+
 # vLLM-Ascend-specific launch flags injected by start_vllm_docker.sh.
 # attention backend for 910C: use the fused/atb attention path. Adjust per image.
 VLLM_ASCEND_ATTENTION_BACKEND="${VLLM_ASCEND_ATTENTION_BACKEND:-atb}"

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh

@@ -74,6 +74,28 @@ engine_stop_server() {
 engine_build_server_args() {
   local tp="$1"
   local dp="$2"
+  # Per-TP overrides (must match start_vllm_docker.sh case $TP).
+  local mem_util="$GPU_MEMORY_UTILIZATION"
+  local max_len="$MAX_MODEL_LEN"
+  local max_seqs="$MAX_NUM_SEQS"
+  case "$tp" in
+    4)
+      mem_util="${TP4_GPU_MEMORY_UTILIZATION:-0.9}"
+      max_len="${TP4_MAX_MODEL_LEN:-32768}"
+      max_seqs="${TP4_MAX_NUM_SEQS:-128}"
+      ;;
+    8)
+      mem_util="${TP8_GPU_MEMORY_UTILIZATION:-0.9}"
+      max_len="${TP8_MAX_MODEL_LEN:-65536}"
+      max_seqs="${TP8_MAX_NUM_SEQS:-256}"
+      ;;
+    16)
+      mem_util="${TP16_GPU_MEMORY_UTILIZATION:-0.9}"
+      max_len="${TP16_MAX_MODEL_LEN:-131072}"
+      max_seqs="${TP16_MAX_NUM_SEQS:-256}"
+      ;;
+  esac
+  log "server args tp=${tp} dp=${dp} gpu_mem_util=${mem_util} max_model_len=${max_len} max_num_seqs=${max_seqs}" >&2
   local -a args=(
     vllm serve "$MODEL_PATH"
     --served-model-name "$SERVED_MODEL_NAME"
@@ -81,9 +103,9 @@ engine_build_server_args() {
     --kv-cache-dtype "$KV_CACHE_DTYPE"
     --block-size "$BLOCK_SIZE"
     --tensor-parallel-size "$tp"
-    --gpu-memory-utilization "$GPU_MEMORY_UTILIZATION"
+    --gpu-memory-utilization "$mem_util"
-    --max-model-len "$MAX_MODEL_LEN"
+    --max-model-len "$max_len"
-    --max-num-seqs "$MAX_NUM_SEQS"
+    --max-num-seqs "$max_seqs"
     --host 0.0.0.0
     --port "$ENGINE_PORT"
   )

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/start_vllm_docker.sh

@@ -18,6 +18,26 @@ SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 # shellcheck source=/dev/null
 source "${SCRIPT_DIR}/config.env"
 
+# Per-TP parameter overrides (synced from glm52 fix 6c81183).
+# DSV4-Flash w8a8 weight per die = ~280GiB / TP; each die has 64GB HBM.
+case "$TP" in
+  4)
+    GPU_MEMORY_UTILIZATION="${TP4_GPU_MEMORY_UTILIZATION:-0.9}"
+    MAX_MODEL_LEN="${TP4_MAX_MODEL_LEN:-32768}"
+    MAX_NUM_SEQS="${TP4_MAX_NUM_SEQS:-128}"
+    ;;
+  8)
+    GPU_MEMORY_UTILIZATION="${TP8_GPU_MEMORY_UTILIZATION:-0.9}"
+    MAX_MODEL_LEN="${TP8_MAX_MODEL_LEN:-65536}"
+    MAX_NUM_SEQS="${TP8_MAX_NUM_SEQS:-256}"
+    ;;
+  16)
+    GPU_MEMORY_UTILIZATION="${TP16_GPU_MEMORY_UTILIZATION:-0.9}"
+    MAX_MODEL_LEN="${TP16_MAX_MODEL_LEN:-131072}"
+    MAX_NUM_SEQS="${TP16_MAX_NUM_SEQS:-256}"
+    ;;
+esac
+
 RUNTIME_BASE="${RUNTIME_BASE:-${SCRIPT_DIR}/runtime}"
 mkdir -p "${RUNTIME_BASE}/logs" "${RUNTIME_BASE}/tmp"