[NV] Add B300 DeepSeek-V4-Pro AgentX with SGLang DSpark6 / 使用 SGLang DSpark6 新增 B300 DeepSeek-V4-Pro AgentX #2624
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 +1,137 @@ | ||
| name: "agg-b300-dep4-mtp-kvoffload" | ||
|
|
||
| model: | ||
| path: "deepseek-v4-pro-0813" | ||
| container: "dynamo-sglang" | ||
| precision: "fp4" | ||
|
|
||
| identity: | ||
| model: | ||
| repo: "deepseek-ai/DeepSeek-V4-Pro-0813" | ||
| container: | ||
| image: "lmsysorg/sglang:nightly-dev-cu13-20260816-4a6dc267" | ||
|
|
||
| dynamo: | ||
| install: true | ||
| hash: "7a4e47ead90ec6b479e397996f6479ce7d251510" | ||
|
|
||
| slurm: | ||
| time_limit: "8:00:00" | ||
|
|
||
| health_check: | ||
| max_attempts: 1440 | ||
| interval_seconds: 10 | ||
|
|
||
| resources: | ||
| gpu_type: "b300" | ||
| gpus_per_node: 4 | ||
| agg_nodes: 1 | ||
| agg_workers: 1 | ||
| gpus_per_agg: 4 | ||
|
|
||
| infra: | ||
| etcd_nats_dedicated_node: true | ||
| nats_max_payload_mb: 32 | ||
|
|
||
| frontend: | ||
| type: dynamo | ||
| nginx_session_affinity: true | ||
| nginx_session_affinity_header: X-Dynamo-Session-ID | ||
| enable_multiple_frontends: false | ||
| env: | ||
| PIP_BREAK_SYSTEM_PACKAGES: "1" | ||
| 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" | ||
|
|
||
| backend: | ||
| type: sglang | ||
| # Let srt-slurm allocate and register the worker-local publisher endpoint; | ||
| # Dynamo needs discovery metadata in addition to the SGLang CLI flag. | ||
| kv_events_config: | ||
| aggregated: true | ||
| aggregated_environment: | ||
| PIP_BREAK_SYSTEM_PACKAGES: "1" | ||
| SGLANG_RAGGED_VERIFY_MODE: "static" | ||
| SGLANG_DEFAULT_THINKING: "1" | ||
| SGLANG_DSV4_REASONING_EFFORT: high | ||
| SGLANG_ENABLE_UNIFIED_RADIX_TREE: "1" | ||
| SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1" | ||
| SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE: "1" | ||
| SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1" | ||
| SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1" | ||
| SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "9216" | ||
| SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "0" | ||
| SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1" | ||
| SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1" | ||
| SGLANG_OPT_USE_JIT_NORM: "1" | ||
| SGLANG_OPT_USE_TOPK_V2: "1" | ||
| sglang_config: | ||
| aggregated: | ||
| served-model-name: "deepseek-ai/DeepSeek-V4-Pro-0813" | ||
| trust-remote-code: true | ||
| enable-metrics: true | ||
| enable-cache-report: true | ||
| # The AgentX corpus contains a small number of traces longer than the | ||
| # runtime KV capacity. Match the existing native B300 launcher: truncate | ||
| # those requests instead of leaving Dynamo streams unresolved. | ||
| allow-auto-truncate: true | ||
| # DSpark draft weights raise the measured minimum viable fraction to | ||
| # 0.9258 on B300; use the same 0.94 fraction as the GB300 TP4 recipe. | ||
| mem-fraction-static: 0.94 | ||
| swa-full-tokens-ratio: 0.02 | ||
| chunked-prefill-size: 32768 | ||
| max-prefill-tokens: 8192 | ||
| max-running-requests: 128 | ||
| cuda-graph-max-bs: 32 | ||
| stream-interval: 20 | ||
| incremental-streaming-output: true | ||
| watchdog-timeout: 1800 | ||
| dp-size: 4 | ||
| tp-size: 4 | ||
| ep-size: 4 | ||
| enable-dp-attention: true | ||
| enable-dp-lm-head: true | ||
| enable-prefill-delayer: true | ||
| prefill-delayer-max-delay-passes: 30 | ||
| moe-dense-tp-size: 1 | ||
| load-balance-method: total_tokens | ||
| moe-a2a-backend: megamoe | ||
| enable-hierarchical-cache: true | ||
| # DEP4's smaller device KV pool otherwise leaves only ~28 GB of host | ||
| # prefix cache per node; ratio 75 raises it to ~700 GB, matching the | ||
| # effective per-session capacity of the validated DEP8 configuration. | ||
| hicache-ratio: 75 | ||
| hicache-write-policy: write_back | ||
| hicache-io-backend: direct | ||
| hicache-mem-layout: page_first_direct | ||
| weight-loader-drop-cache-after-load: true | ||
| speculative-algorithm: DSPARK | ||
| speculative-dspark-block-size: 6 | ||
| speculative-num-steps: 1 | ||
| speculative-eagle-topk: 1 | ||
| speculative-num-draft-tokens: 7 | ||
|
|
||
| sbatch_directives: | ||
| mem: "0" | ||
| cpus-per-task: "144" | ||
|
|
||
| srun_options: | ||
| mem: "0" | ||
| container-remap-root: "" | ||
|
|
||
| benchmark: | ||
| type: custom | ||
| command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh | ||
| env: | ||
| INFMAX_CONTAINER_WORKSPACE: /infmax-workspace | ||
| RESULT_DIR: /logs/agentic | ||
| PORT: "8000" | ||
| IS_MULTINODE: "false" | ||
| TP: "4" | ||
| AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true" | ||
| AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0" | ||
| AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:" | ||
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 +1,136 @@ | ||||||||
| name: "agg-b300-dep8-mtp-kvoffload" | ||||||||
|
|
||||||||
| model: | ||||||||
| path: "deepseek-v4-pro-0813" | ||||||||
| container: "dynamo-sglang" | ||||||||
| precision: "fp4" | ||||||||
|
|
||||||||
| identity: | ||||||||
| model: | ||||||||
| repo: "deepseek-ai/DeepSeek-V4-Pro-0813" | ||||||||
| container: | ||||||||
| image: "lmsysorg/sglang:nightly-dev-cu13-20260816-4a6dc267" | ||||||||
|
|
||||||||
| dynamo: | ||||||||
| install: true | ||||||||
| hash: "7a4e47ead90ec6b479e397996f6479ce7d251510" | ||||||||
|
|
||||||||
| slurm: | ||||||||
| time_limit: "8:00:00" | ||||||||
|
|
||||||||
| health_check: | ||||||||
| max_attempts: 1440 | ||||||||
| interval_seconds: 10 | ||||||||
|
|
||||||||
| resources: | ||||||||
| gpu_type: "b300" | ||||||||
| gpus_per_node: 8 | ||||||||
| agg_nodes: 1 | ||||||||
| agg_workers: 1 | ||||||||
| gpus_per_agg: 8 | ||||||||
|
|
||||||||
| infra: | ||||||||
| etcd_nats_dedicated_node: true | ||||||||
| nats_max_payload_mb: 32 | ||||||||
|
|
||||||||
| frontend: | ||||||||
| type: dynamo | ||||||||
| nginx_session_affinity: true | ||||||||
| nginx_session_affinity_header: X-Dynamo-Session-ID | ||||||||
| enable_multiple_frontends: false | ||||||||
| env: | ||||||||
| PIP_BREAK_SYSTEM_PACKAGES: "1" | ||||||||
| 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" | ||||||||
|
|
||||||||
| backend: | ||||||||
| type: sglang | ||||||||
| # Let srt-slurm allocate and register the worker-local publisher endpoint; | ||||||||
| # Dynamo needs discovery metadata in addition to the SGLang CLI flag. | ||||||||
| kv_events_config: | ||||||||
| aggregated: true | ||||||||
| aggregated_environment: | ||||||||
| PIP_BREAK_SYSTEM_PACKAGES: "1" | ||||||||
| SGLANG_RAGGED_VERIFY_MODE: "static" | ||||||||
| SGLANG_DEFAULT_THINKING: "1" | ||||||||
|
Contributor
There was a problem hiding this comment. 🟡 WARNING: Same as DEP4 — missing Fix:
Suggested change
|
||||||||
| SGLANG_DSV4_REASONING_EFFORT: high | ||||||||
| SGLANG_ENABLE_UNIFIED_RADIX_TREE: "1" | ||||||||
| SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1" | ||||||||
| SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE: "1" | ||||||||
| SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1" | ||||||||
| SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1" | ||||||||
| SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "9216" | ||||||||
| SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "0" | ||||||||
| SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1" | ||||||||
| SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1" | ||||||||
| SGLANG_OPT_USE_JIT_NORM: "1" | ||||||||
| SGLANG_OPT_USE_TOPK_V2: "1" | ||||||||
| sglang_config: | ||||||||
| aggregated: | ||||||||
| served-model-name: "deepseek-ai/DeepSeek-V4-Pro-0813" | ||||||||
| trust-remote-code: true | ||||||||
| enable-metrics: true | ||||||||
| enable-cache-report: true | ||||||||
| # The AgentX corpus contains a small number of traces longer than the | ||||||||
| # runtime KV capacity. Match the existing native B300 launcher: truncate | ||||||||
| # those requests instead of leaving Dynamo streams unresolved. | ||||||||
| allow-auto-truncate: true | ||||||||
| # Leave activation headroom for 8K/rank chunked prefill; 0.90 OOMs | ||||||||
| # under the c128 AgentX warmup while 0.85 completes the full run. | ||||||||
| mem-fraction-static: 0.85 | ||||||||
| swa-full-tokens-ratio: 0.02 | ||||||||
| chunked-prefill-size: 65536 | ||||||||
| max-prefill-tokens: 8192 | ||||||||
| max-running-requests: 1152 | ||||||||
| cuda-graph-max-bs: 144 | ||||||||
| stream-interval: 20 | ||||||||
| incremental-streaming-output: true | ||||||||
| watchdog-timeout: 1800 | ||||||||
| dp-size: 8 | ||||||||
| tp-size: 8 | ||||||||
| ep-size: 8 | ||||||||
| enable-dp-attention: true | ||||||||
| enable-dp-lm-head: true | ||||||||
| enable-prefill-delayer: true | ||||||||
| prefill-delayer-max-delay-passes: 30 | ||||||||
| moe-dense-tp-size: 1 | ||||||||
| load-balance-method: total_tokens | ||||||||
| moe-a2a-backend: megamoe | ||||||||
| enable-hierarchical-cache: true | ||||||||
| # Keep the explicit host KV pool within the x86 B300 node's 3 TB DRAM | ||||||||
| # budget; ratio 2.0 OOMs during startup before the benchmark can run. | ||||||||
| hicache-ratio: 1.5 | ||||||||
| hicache-write-policy: write_back | ||||||||
| hicache-io-backend: direct | ||||||||
| hicache-mem-layout: page_first_direct | ||||||||
| weight-loader-drop-cache-after-load: true | ||||||||
| speculative-algorithm: DSPARK | ||||||||
| speculative-dspark-block-size: 6 | ||||||||
| speculative-num-steps: 1 | ||||||||
| speculative-eagle-topk: 1 | ||||||||
| speculative-num-draft-tokens: 7 | ||||||||
|
|
||||||||
| sbatch_directives: | ||||||||
| mem: "0" | ||||||||
| cpus-per-task: "144" | ||||||||
|
|
||||||||
| srun_options: | ||||||||
| mem: "0" | ||||||||
| container-remap-root: "" | ||||||||
|
|
||||||||
| benchmark: | ||||||||
| type: custom | ||||||||
| command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh | ||||||||
| env: | ||||||||
| INFMAX_CONTAINER_WORKSPACE: /infmax-workspace | ||||||||
| RESULT_DIR: /logs/agentic | ||||||||
| PORT: "8000" | ||||||||
| IS_MULTINODE: "false" | ||||||||
| TP: "8" | ||||||||
| AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true" | ||||||||
| AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0" | ||||||||
| AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:" | ||||||||
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
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,19 @@ | ||
| deepseek-v4-pro-0813: | ||
| thinking_on: | ||
| 1: 1.84 | ||
| 2: 2.51 | ||
| 3: 3.01 | ||
| 4: 3.36 | ||
| 5: 3.61 | ||
| 6: 3.77 | ||
| 7: 3.73 | ||
| 8: 3.47 | ||
| thinking_off: | ||
| 1: 1.94 | ||
| 2: 2.81 | ||
| 3: 3.58 | ||
| 4: 4.17 | ||
| 5: 4.74 | ||
| 6: 5.12 | ||
| 7: 5.24 | ||
| 8: 4.95 |
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.
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.
🟡 WARNING: Missing
SGLANG_DSV4_REASONING_EFFORT: highinaggregated_environment.Why it matters: Every existing DSV4 AgentX recipe sets it in each worker env block (GB300
agg-gb300-tp4/tp8-mtp-lowlatency.yaml, all fourdisagg-gb300-*-mtp-kvoffload.yaml, andagg-h200-tp8-mtp-kvoffload.yaml). Without it the model runs at the default reasoning effort, which changes generation length/profile — so the B300 curve isn't comparable with the GB300/H200 DSV4 AgentX results this PR claims parity with, and the golden acceptance length 3.82 (measured at efforthigh) no longer matches what eval-only real-verify runs will observe.Fix:
(Same fix needed in
agg-b300-dep8-mtp-kvoffload.yamlline 60.)