feat(910c/glm52): 并行配置改为 4 4 / 8 2 / 16 1 对标H20的 2 4 / 4 2 / 8 1

A3 910C 有16 dies(8卡x2die),H20有8卡。为公平对比,TP按卡数等效:
  A3 TP=4  DP=4 (4die/副本x4) == H20 TP=2 DP=4 (2卡/副本x4)
  A3 TP=8  DP=2 (8die/副本x2) == H20 TP=4 DP=2 (4卡/副本x2)
  A3 TP=16 DP=1 (16die/副本x1) == H20 TP=8 DP=1 (8卡/副本x1)

与dsv4实验(a65849b)保持一致的配置思路。

新增TP=4 per-TP参数覆盖(expert-parallel下专家分4份+dense复制,
KV cache极紧): gpu_mem=0.97, max_model_len=4096, max_num_seqs=32
TP=4可能OOM,若发生会自动记录并跳过。

config.env: PARALLEL_CONFIGS 8 1/16 1 -> 4 4/8 2/16 1; 加TP4_变量
start_vllm_docker.sh: case $TP 加 TP=4 分支
run_adaptive_concurrency_add16.sh: case $tp 加 TP=4 分支

experiments/910c/glm52_910c_vllm_tp_dp_matrix/config.env

@@ -47,9 +47,12 @@ 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"
-# Each Ascend910 card has 2 dies; TP addresses dies. TP=8 uses 8 dies (4 cards),
-# TP=16 uses all 16 dies (8 cards). Only these two configs are tested.
-# Override via PARALLEL_CONFIGS_STR="8,1 16,1" (space-separated).
+# A3 910C has 16 dies (8 cards x 2 dies/card). TP/DP address dies, so TP*DP=16
+# uses all dies. The mapping to H20 (8 cards) for fair comparison is:
+#   A3 TP=4  DP=4 -> 4 dies/replica x 4 replicas  == H20 TP=2 DP=4 (2 cards/replica)
+#   A3 TP=8  DP=2 -> 8 dies/replica x 2 replicas  == H20 TP=4 DP=2 (4 cards/replica)
+#   A3 TP=16 DP=1 -> 16 dies/replica x 1 replica   == H20 TP=8 DP=1 (8 cards/replica)
+# Override via PARALLEL_CONFIGS_STR="4,4 8,2 16,1" (space-separated).
 if [[ -n "${PARALLEL_CONFIGS_STR:-}" ]]; then
   declare -a PARALLEL_CONFIGS=()
   for pair in $PARALLEL_CONFIGS_STR; do
@@ -57,7 +60,8 @@ if [[ -n "${PARALLEL_CONFIGS_STR:-}" ]]; then
   done
 else
   declare -a PARALLEL_CONFIGS=(
-    "8 1"
+    "4 4"
+    "8 2"
     "16 1"
   )
 fi
@@ -74,6 +78,22 @@ 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 $TP).
+# GLM-5.2-w4a8c8 ~391 GiB total; with expert-parallel expert weights are sharded
+# across TP dies but dense (attn) weights are replicated. Each die has 64GB HBM.
+#   TP=4:  experts/4 + full dense; very tight KV cache -> cap context heavily
+#   TP=8:  experts/8 + full dense; tight KV cache (verified: max_len=16384 fits)
+#   TP=16: experts/16 + full dense; ample KV cache (full 128K context)
+TP4_GPU_MEMORY_UTILIZATION="${TP4_GPU_MEMORY_UTILIZATION:-0.97}"
+TP4_MAX_MODEL_LEN="${TP4_MAX_MODEL_LEN:-4096}"
+TP4_MAX_NUM_SEQS="${TP4_MAX_NUM_SEQS:-32}"
+TP8_GPU_MEMORY_UTILIZATION="${TP8_GPU_MEMORY_UTILIZATION:-0.95}"
+TP8_MAX_MODEL_LEN="${TP8_MAX_MODEL_LEN:-16384}"
+TP8_MAX_NUM_SEQS="${TP8_MAX_NUM_SEQS:-64}"
+TP16_GPU_MEMORY_UTILIZATION="${TP16_GPU_MEMORY_UTILIZATION:-0.92}"
+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/glm52_910c_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh

@@ -80,6 +80,11 @@ engine_build_server_args() {
   local max_len="$MAX_MODEL_LEN"
   local max_seqs="$MAX_NUM_SEQS"
   case "$tp" in
+    4)
+      mem_util="${TP4_GPU_MEMORY_UTILIZATION:-0.97}"
+      max_len="${TP4_MAX_MODEL_LEN:-4096}"
+      max_seqs="${TP4_MAX_NUM_SEQS:-32}"
+      ;;
     8)
       mem_util="${TP8_GPU_MEMORY_UTILIZATION:-0.95}"
       max_len="${TP8_MAX_MODEL_LEN:-16384}"

experiments/910c/glm52_910c_vllm_tp_dp_matrix/start_vllm_docker.sh

@@ -34,6 +34,11 @@ docker rm -f "$NAME" >/dev/null 2>&1 || true
 # Per-TP overrides: TP=8 has tight KV cache on 64GB/die (cap context/seqs),
 # TP=16 uses all 16 dies and can serve the full 128K context.
 case "$TP" in
+  4)
+    GPU_MEMORY_UTILIZATION="${TP4_GPU_MEMORY_UTILIZATION:-0.97}"
+    MAX_MODEL_LEN="${TP4_MAX_MODEL_LEN:-4096}"
+    MAX_NUM_SEQS="${TP4_MAX_NUM_SEQS:-32}"
+    ;;
   8)
     GPU_MEMORY_UTILIZATION="${TP8_GPU_MEMORY_UTILIZATION:-0.95}"
     MAX_MODEL_LEN="${TP8_MAX_MODEL_LEN:-16384}"