fix(dsv4): align launch params with official A3 tutorial (fixes DP die allocation)

The previous params caused all 4 DP replicas to bind to the same 4 dies
(die 0-3), leaving 12 dies idle. Root cause was a combination of missing
official params + extra non-official params that interfered with DP
worker device placement.

Verified: with aligned params, TP=4 DP=4 correctly distributes 16 workers
across all 16 dies (8 cards x 2 dies), each at ~57 GB HBM (91% util).

Changes (align to docs.vllm.ai A3 tutorial):
- Add --max-num-batched-tokens 10240 (was missing; affects DP scheduling)
- Add --api-server-count 1 (was missing; without it vllm spawns N API
  servers for N DP ranks, disturbing device assignment)
- Remove --kv-cache-dtype fp8 (official uses default bfloat16)
- Remove --trust-remote-code (official doesn't use it for DSV4)
- Remove enable_dsa_cp from additional-config (official doesn't have it)
- max-model-len: per-TP caps (32768/65536/131072) -> 1048576 for all TPs
  (official uses full 1M; the caps were over-cautious)
- max-num-seqs: per-TP (128/256/256) -> 64 for all (official value)

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/config.env

@@ -79,10 +79,14 @@ fi
 #     matrix sweep (extend to 1M via matrix.json once TP=8 is verified).
 #   - gpu-memory-utilization maps to NPU HBM fraction on vllm-ascend (0.9).
 GPU_MEMORY_UTILIZATION="${GPU_MEMORY_UTILIZATION:-0.9}"
-KV_CACHE_DTYPE="${KV_CACHE_DTYPE:-fp8}"
+KV_CACHE_DTYPE="${KV_CACHE_DTYPE:-}"
 BLOCK_SIZE="${BLOCK_SIZE:-128}"
-MAX_MODEL_LEN="${MAX_MODEL_LEN:-131072}"
-MAX_NUM_SEQS="${MAX_NUM_SEQS:-256}"
+MAX_MODEL_LEN="${MAX_MODEL_LEN:-1048576}"
+MAX_NUM_SEQS="${MAX_NUM_SEQS:-64}"
+# Max tokens per batch (official A3 tutorial: 10240).
+MAX_NUM_BATCHED_TOKENS="${MAX_NUM_BATCHED_TOKENS:-10240}"
+# API server count (official: 1).
+API_SERVER_COUNT="${API_SERVER_COUNT:-1}"
 
 # Per-TP parameter overrides (applied in start_vllm_docker.sh via case $TP).
 # DSV4-Flash w8a8 weight per die = ~280GiB / TP. Each die has 64GB HBM.
@@ -90,14 +94,14 @@ MAX_NUM_SEQS="${MAX_NUM_SEQS:-256}"
 #   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="${TP4_MAX_NUM_SEQS:-128}"
+TP4_MAX_MODEL_LEN="${TP4_MAX_MODEL_LEN:-1048576}"
+TP4_MAX_NUM_SEQS="${TP4_MAX_NUM_SEQS:-64}"
 TP8_GPU_MEMORY_UTILIZATION="${TP8_GPU_MEMORY_UTILIZATION:-0.9}"
-TP8_MAX_MODEL_LEN="${TP8_MAX_MODEL_LEN:-65536}"
-TP8_MAX_NUM_SEQS="${TP8_MAX_NUM_SEQS:-256}"
+TP8_MAX_MODEL_LEN="${TP8_MAX_MODEL_LEN:-1048576}"
+TP8_MAX_NUM_SEQS="${TP8_MAX_NUM_SEQS:-64}"
 TP16_GPU_MEMORY_UTILIZATION="${TP16_GPU_MEMORY_UTILIZATION:-0.9}"
-TP16_MAX_MODEL_LEN="${TP16_MAX_MODEL_LEN:-131072}"
-TP16_MAX_NUM_SEQS="${TP16_MAX_NUM_SEQS:-256}"
+TP16_MAX_MODEL_LEN="${TP16_MAX_MODEL_LEN:-1048576}"
+TP16_MAX_NUM_SEQS="${TP16_MAX_NUM_SEQS:-64}"
 
 # vLLM-Ascend-specific launch flags injected by start_vllm_docker.sh.
 # attention backend for 910C: use the fused/atb attention path. Adjust per image.
@@ -123,7 +127,7 @@ DSV4_SAFETENSORS_LOAD_STRATEGY="${DSV4_SAFETENSORS_LOAD_STRATEGY:-prefetch}"
 # in the launcher to avoid shell mangling.
 DSV4_SPEC_CONFIG="${DSV4_SPEC_CONFIG:-{\"num_speculative_tokens\": 1, \"method\": \"mtp\", \"enforce_eager\": true}}"
 DSV4_COMPILATION_CONFIG="${DSV4_COMPILATION_CONFIG:-{\"cudagraph_mode\": \"FULL_DECODE_ONLY\"}}"
-DSV4_ADDITIONAL_CONFIG="${DSV4_ADDITIONAL_CONFIG:-{\"ascend_compilation_config\":{\"enable_npugraph_ex\":true,\"enable_static_kernel\":false},\"enable_cpu_binding\": true,\"enable_dsa_cp\": true,\"multistream_overlap_shared_expert\":true}}"
+DSV4_ADDITIONAL_CONFIG="${DSV4_ADDITIONAL_CONFIG:-{\"ascend_compilation_config\":{\"enable_npugraph_ex\":true,\"enable_static_kernel\":false},\"enable_cpu_binding\": true,\"multistream_overlap_shared_expert\":true}}"
 DSV4_ENABLE_EXPERT_PARALLEL="${DSV4_ENABLE_EXPERT_PARALLEL:-1}"
 DSV4_ENABLE_ASYNC_SCHEDULING="${DSV4_ENABLE_ASYNC_SCHEDULING:-1}"
 

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh

@@ -99,13 +99,13 @@ engine_build_server_args() {
   local -a args=(
     vllm serve "$MODEL_PATH"
     --served-model-name "$SERVED_MODEL_NAME"
-    --trust-remote-code
-    --kv-cache-dtype "$KV_CACHE_DTYPE"
-    --block-size "$BLOCK_SIZE"
+        --block-size "$BLOCK_SIZE"
     --tensor-parallel-size "$tp"
     --gpu-memory-utilization "$mem_util"
     --max-model-len "$max_len"
     --max-num-seqs "$max_seqs"
+  --max-num-batched-tokens "$MAX_NUM_BATCHED_TOKENS"
+  --api-server-count "$API_SERVER_COUNT"
     --host 0.0.0.0
     --port "$ENGINE_PORT"
   )

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/start_vllm_docker.sh

@@ -59,13 +59,13 @@ docker rm -f "$NAME" >/dev/null 2>&1 || true
 SERVER_ARGS=(
   "$MODEL_PATH"
   --served-model-name "$SERVED_MODEL_NAME"
-  --trust-remote-code
-  --kv-cache-dtype "$KV_CACHE_DTYPE"
   --block-size "$BLOCK_SIZE"
   --tensor-parallel-size "$TP"
   --gpu-memory-utilization "$GPU_MEMORY_UTILIZATION"
   --max-model-len "$MAX_MODEL_LEN"
   --max-num-seqs "$MAX_NUM_SEQS"
+  --max-num-batched-tokens "$MAX_NUM_BATCHED_TOKENS"
+  --api-server-count "$API_SERVER_COUNT"
   --host 0.0.0.0
   --port "$PORT"
 )