Repository navigation
[AgentX] MiniMax-M3.1-Flash GB200 SGLang: update config / [AgentX] MiniMax-M3.1-Flash GB200 SGLang：更新配置 #4695
+516
−0
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,204 @@ | ||
| schema: 2 | ||
|
|
||
| base: | ||
| name: minimaxm31flash-fp4-gb200-dynamo-sglang-agentx-disagg | ||
| identity: | ||
| model: {} | ||
| frameworks: | ||
| dynamo: "1.7.0.dev20261030" | ||
| sglang: "0.0.0.dev1+ge3718813e" | ||
| dynamo: | ||
| install: false | ||
| request_plane: tcp | ||
| event_plane: zmq | ||
| environment: | ||
| HF_HUB_CACHE: /hf_hub_cache | ||
| # Tachometer's per-node exporters slow decode steps and add host memory; | ||
| # InferenceX does not read its output. | ||
| observability: | ||
| tachometer: | ||
| enabled: false | ||
| slurm: | ||
| time_limit: "6:00:00" | ||
| health_check: | ||
| max_attempts: 1440 | ||
| interval_seconds: 10 | ||
| resources: | ||
| gpu_type: gb200 | ||
| gpus_per_node: 4 | ||
| services: | ||
| - name: etcd | ||
| type: etcd | ||
| placement: | ||
| node: infra | ||
| # Disables Dynamo's per-request lifecycle span export; InferenceX does not read it. | ||
| - name: otel-collector | ||
| type: otel-collector | ||
| enabled: false | ||
| # Keeps tokenize, route and stream CPU work off the engine trays. | ||
| frontend: | ||
| type: dynamo | ||
| enable_multiple_frontends: false | ||
| placement: | ||
| node: dedicated | ||
| env: | ||
| DYN_ROUTER_TEMPERATURE: "0" | ||
| DYN_ROUTER_QUEUE_THRESHOLD: None | ||
| DYN_ROUTER_ACTIVE_REQUEST_EXPIRY_SECS: "3600" | ||
| DYN_TCP_CONNECT_TIMEOUT: "120" | ||
| # The default frontend tokenizer disables its prefix cache for this model, | ||
| # so every turn would re-tokenize the full prompt. | ||
| DYN_TOKENIZER: "fastokens" | ||
| DYN_TOKENIZER_CACHE: "1" | ||
| DYN_TOKENIZER_CACHE_BYTES: "8000000000" | ||
| args: | ||
| router-mode: kv | ||
| router-session-affinity-ttl-secs: "14400" | ||
| active-decode-blocks-threshold: "None" | ||
| active-prefill-tokens-threshold: "None" | ||
| active-prefill-tokens-threshold-frac: "None" | ||
| engine: sglang | ||
| roles: | ||
| prefill: | ||
| nodes: 2 | ||
| workers: 2 | ||
| gpus: 4 | ||
| env: | ||
| PYTHONUNBUFFERED: "1" | ||
| PYTHONNOUSERSITE: "1" | ||
| FMHA_SM100_ALLOW_JIT: "1" | ||
| SGLANG_TIMEOUT_KEEP_ALIVE: "1800" | ||
| SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "3600" | ||
| SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "3600" | ||
| SGLANG_MOONCAKE_CUSTOM_MEM_POOL: "True" | ||
| SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "32768" | ||
| NCCL_MNNVL_ENABLE: "1" | ||
| NCCL_CUMEM_ENABLE: "1" | ||
| MC_FORCE_MNNVL: "1" | ||
| MC_ENABLE_DEST_DEVICE_AFFINITY: "1" | ||
| # Some nodes list a BMC interface with the same address on every node first. | ||
| MC_TCP_BIND_ADDRESS: "{node}" | ||
| WITH_NVIDIA_PEERMEM: "0" | ||
| # Dynamo 1.7 backend admission rejects turns with 529 once a DP rank is full; | ||
| # the SGLang scheduler queues them instead. | ||
| DYN_BACKEND_ADMISSION_ENGINE_REQUEST_LIMIT: "10000" | ||
| DYN_BACKEND_ADMISSION_ENGINE_REQUEST_WAITING_LIMIT: "10000" | ||
| DYN_BACKEND_ADMISSION_REQUEST_QUEUE_LIMIT: "10000" | ||
| args: | ||
| served-model-name: MiniMaxAI/MiniMax-M3.1-Flash | ||
| trust-remote-code: true | ||
| reasoning-parser: minimax-m3 | ||
| tool-call-parser: minimax-m3 | ||
| tensor-parallel-size: 4 | ||
| data-parallel-size: 4 | ||
| expert-parallel-size: 4 | ||
| enable-dp-attention: true | ||
| enable-dp-lm-head: true | ||
| # flashinfer_trtllm_routed needs attention TP to match EP, which attention DP breaks. | ||
| moe-a2a-backend: megamoe | ||
| disaggregation-transfer-backend: mooncake | ||
| chunked-prefill-size: 8192 | ||
| # Shortest uncached prefill first: short cached turns stop queueing | ||
| # behind long uncached prompts, which cuts the TTFT tail. | ||
| schedule-policy: shortest-prefill-first | ||
| # 64 per DP rank. | ||
| max-running-requests: 256 | ||
| mem-fraction-static: 0.85 | ||
| enable-hierarchical-cache: true | ||
| # 180 GB per rank: 720 of the node's ~900 GB host allowance, leaving | ||
| # headroom for the engine, Mooncake and the OS. | ||
| hicache-size: 180 | ||
| hicache-write-policy: write_back | ||
| speculative-algorithm: DSPARK | ||
| speculative-dspark-block-size: 7 | ||
| speculative-use-block-verification: true | ||
| speculative-draft-kv-cache-dtype: fp8_e4m3 | ||
| # A session's turns stay on one prefill worker. | ||
| router-mode: kv | ||
| router-session-affinity-ttl-secs: "14400" | ||
| watchdog-timeout: 3600 | ||
| enable-metrics: true | ||
| enable-cache-report: true | ||
| kv_events: true | ||
| decode: | ||
| nodes: 2 | ||
| workers: 1 | ||
| gpus: 8 | ||
| env: | ||
| PYTHONUNBUFFERED: "1" | ||
| PYTHONNOUSERSITE: "1" | ||
| FMHA_SM100_ALLOW_JIT: "1" | ||
| SGLANG_TIMEOUT_KEEP_ALIVE: "1800" | ||
| DYN_SGL_GATEWAY_WORKERS: "8" | ||
| SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "3600" | ||
| SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "3600" | ||
| SGLANG_MOONCAKE_CUSTOM_MEM_POOL: "True" | ||
| SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "16384" | ||
| NCCL_MNNVL_ENABLE: "1" | ||
| NCCL_CUMEM_ENABLE: "1" | ||
| MC_FORCE_MNNVL: "1" | ||
| MC_ENABLE_DEST_DEVICE_AFFINITY: "1" | ||
| MC_TCP_BIND_ADDRESS: "{node}" | ||
| WITH_NVIDIA_PEERMEM: "0" | ||
| DYN_BACKEND_ADMISSION_ENGINE_REQUEST_LIMIT: "10000" | ||
| DYN_BACKEND_ADMISSION_ENGINE_REQUEST_WAITING_LIMIT: "10000" | ||
| DYN_BACKEND_ADMISSION_REQUEST_QUEUE_LIMIT: "10000" | ||
| args: | ||
| served-model-name: MiniMaxAI/MiniMax-M3.1-Flash | ||
| trust-remote-code: true | ||
| reasoning-parser: minimax-m3 | ||
| tool-call-parser: minimax-m3 | ||
| tensor-parallel-size: 8 | ||
| data-parallel-size: 8 | ||
| expert-parallel-size: 8 | ||
| enable-dp-attention: true | ||
| enable-dp-lm-head: true | ||
| moe-a2a-backend: megamoe | ||
| disaggregation-transfer-backend: mooncake | ||
| # The frontend picks the decode DP rank; least-loaded sends each request to | ||
| # the rank with the fewest in flight, where round-robin left the busiest rank | ||
| # at 1.45x the mean and every DP rank waits for it each step. | ||
| router-mode: least-loaded | ||
| load-balance-method: total_tokens | ||
| # 48 per DP rank. | ||
| max-running-requests: 384 | ||
| cuda-graph-max-bs-decode: 48 | ||
| mem-fraction-static: 0.85 | ||
| speculative-algorithm: DSPARK | ||
| speculative-dspark-block-size: 7 | ||
| speculative-use-block-verification: true | ||
| speculative-draft-kv-cache-dtype: fp8_e4m3 | ||
| watchdog-timeout: 3600 | ||
| enable-metrics: true | ||
| kv_events: false | ||
| sbatch_directives: | ||
| mem: "0" | ||
| exclusive: "" | ||
| srun_options: | ||
| container-remap-root: "" | ||
| mem: "0" | ||
| benchmark: | ||
| placement: | ||
| node: dedicated | ||
| env: | ||
| AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true" | ||
| AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0" | ||
| AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:" | ||
|
|
||
| override_2p_dep4_dep8: | ||
| name: minimaxm31flash-fp4-gb200-dynamo-sglang-agentx-disagg-2p-dep4-dep8 | ||
|
|
||
| override_3p_dep4_dep16: | ||
| name: minimaxm31flash-fp4-gb200-dynamo-sglang-agentx-disagg-3p-dep4-dep16 | ||
| roles: | ||
| prefill: | ||
| nodes: 3 | ||
| workers: 3 | ||
| decode: | ||
| nodes: 4 | ||
| gpus: 16 | ||
| args: | ||
| tensor-parallel-size: 16 | ||
| data-parallel-size: 16 | ||
| expert-parallel-size: 16 | ||
| max-running-requests: 768 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
🟡 (optional) Published results for this config will claim 901 GB of host KV per node, while the recipe only pins 720 GB.
dram-utilization: 1.0makes matrix generation record the whole gb200-nv node share (860_160 MiB, about 901 GB) astotal-cpu-dram-gb. The prefill recipe pinshicache-size: 180on 4 ranks, which is 720 GB. Fix: setdram-utilization: 0.80, which resolves to 721 GB. The single-node siblingminimaxm31flash-fp4-gb200-sglang-agentic-dsparkuses that value for the same 180 GB-per-rank pool.Why this was flagged
agentic_dram_offload_gb (infx/matrix/generate.py:475-518) computes the prefill share as available-cpu-dram-mib 860_160 (runners.yaml:572) × utilization × 4/4 GPUs. With 1.0 that comes to 901 GB. The value reaches TOTAL_CPU_DRAM_GB through infx/srt_slurm/generate.py:59. It is published as
allocated_cpu_dram_gb(infx/results/agentic/init.py:205). The decode side also gets 901 GB asdecode_allocated_cpu_dram_gb, even though decode runs no HiCache. The recipe really allocates 180 GB × 4 ranks = 720 GB per prefill node; its own comment says "720 of the node's ~900 GB host allowance". The same-cluster, same-pool sibling uses dram-utilization 0.80. Its recipe records TOTAL_CPU_DRAM_GB '721' (benchmarks/single_node/.../minimaxm31flash/sglang/gb200-fp4-mtp/agentic.yaml:148). So results with identical host pools would report different DRAM budgets, and this one overstates its pool by about 25%. Nothing checks the recorded value against hicache-size, because the recipe never reads TOTAL_CPU_DRAM_GB.Verification: nit. This happens on every DRAM-offload point of the new config. In
agentic_dram_offload_gb,share_gbreturns 860160 MiB × 1.0 × 4/4, which is about 901 GB.infx/results/agentic/__init__.py:205publishes this asallocated_cpu_dram_gb. The recipe really pinshicache-size: 180on 4 ranks per prefill node, which is 720 GB. Nothing fails at runtime.