-
Notifications
You must be signed in to change notification settings - Fork 176
[Do Not Merge] Add GB300 DSV4 Dynamo vLLM MTP recipes #1535
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Draft
hjjq
wants to merge
8
commits into
main
Choose a base branch
from
hjjq/gb300-mtp
base: main
Could not load branches
Branch not found: {{ refName }}
Loading
Could not load tags
Nothing to show
Loading
Are you sure you want to change the base?
Some commits from the old base branch may be removed from the timeline,
and old review comments may become outdated.
Draft
Changes from all commits
Commits
Show all changes
8 commits
Select commit
Hold shift + click to select a range
0e1f762
add recipes
hjjq 8ca7f0e
Merge remote-tracking branch 'origin/main' into hjjq/gb300-mtp
hjjq b190811
remove identity block
hjjq 530b5bc
Merge branch 'main' into hjjq/gb300-mtp
Ankur-singh dec6b36
gb300 runner: pin dsv4-vllm to aflowers/vllm-gb200-v0.20.0
hjjq f2265a7
Merge remote-tracking branch 'origin/main' into hjjq/gb300-mtp
hjjq 2f2b06a
trigger ci
hjjq 2c40df8
Merge remote-tracking branch 'origin/main' into hjjq/gb300-mtp
hjjq File filter
Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
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
120 changes: 120 additions & 0 deletions
120
...ode/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-1p1d-dep4-dep8-megamoe-mtp2.yaml
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 +1,120 @@ | ||
| name: "svf-vllm-disagg-gb300-1p1d-dep4-dep8-megamoe-mtp2" | ||
|
|
||
| model: | ||
| path: "deepseek-v4-pro" | ||
| container: "vllm/vllm-openai:v0.21.0-ubuntu2404" | ||
| precision: "fp4" | ||
|
|
||
| dynamo: | ||
| install: true | ||
| wheel: "1.2.0.dev20260426" | ||
|
|
||
| setup_script: vllm-container-deps.sh | ||
|
|
||
| slurm: | ||
| time_limit: "8:00:00" | ||
|
|
||
| health_check: | ||
| max_attempts: 1440 | ||
| interval_seconds: 10 | ||
|
|
||
| resources: | ||
| gpu_type: "gb300" | ||
| gpus_per_node: 4 | ||
| prefill_nodes: 1 | ||
| decode_nodes: 2 | ||
| prefill_workers: 1 | ||
| decode_workers: 1 | ||
| gpus_per_prefill: 4 | ||
| gpus_per_decode: 8 | ||
|
|
||
| infra: | ||
| etcd_nats_dedicated_node: true | ||
|
|
||
| frontend: | ||
| type: dynamo | ||
| enable_multiple_frontends: false | ||
|
|
||
| backend: | ||
| type: vllm | ||
| connector: null | ||
| prefill_environment: | ||
| VLLM_ENGINE_READY_TIMEOUT_S: "3600" | ||
| TILELANG_CLEANUP_TEMP_FILES: "1" | ||
| VLLM_USE_NCCL_SYMM_MEM: "1" | ||
| NCCL_CUMEM_ENABLE: "1" | ||
| NCCL_MNNVL_ENABLE: "1" | ||
| NCCL_NVLS_ENABLE: "1" | ||
| TORCH_SYMMMEM: "NVSHMEM" | ||
| decode_environment: | ||
| VLLM_ENGINE_READY_TIMEOUT_S: "3600" | ||
| TILELANG_CLEANUP_TEMP_FILES: "1" | ||
| VLLM_USE_NCCL_SYMM_MEM: "1" | ||
| NCCL_CUMEM_ENABLE: "1" | ||
| NCCL_MNNVL_ENABLE: "1" | ||
| NCCL_NVLS_ENABLE: "1" | ||
| TORCH_SYMMMEM: "NVSHMEM" | ||
|
|
||
| vllm_config: | ||
| prefill: | ||
| kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' | ||
| served-model-name: "deepseek-ai/DeepSeek-V4-Pro" | ||
| kv-cache-dtype: "fp8" | ||
| tensor-parallel-size: 1 | ||
| pipeline-parallel-size: 1 | ||
| data-parallel-size: 4 | ||
| data-parallel-rpc-port: 13345 | ||
| enable-expert-parallel: true | ||
| enable-ep-weight-filter: true | ||
| moe-backend: deep_gemm_mega_moe | ||
| enforce-eager: true | ||
| attention-config: '{"use_fp4_indexer_cache":true}' | ||
| max-model-len: 16384 | ||
| max-num-seqs: 256 | ||
| max-num-batched-tokens: 16384 | ||
| trust-remote-code: true | ||
| no-enable-prefix-caching: true | ||
| no-enable-flashinfer-autotune: true | ||
| no-async-scheduling: true | ||
| block-size: 256 | ||
| gpu-memory-utilization: 0.9 | ||
| no-disable-hybrid-kv-cache-manager: true | ||
| enable-sleep-mode: true | ||
| tokenizer-mode: deepseek_v4 | ||
| decode: | ||
| kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' | ||
| served-model-name: "deepseek-ai/DeepSeek-V4-Pro" | ||
| kv-cache-dtype: "fp8" | ||
| tensor-parallel-size: 1 | ||
| pipeline-parallel-size: 1 | ||
| data-parallel-size: 8 | ||
| data-parallel-rpc-port: 13345 | ||
| enable-expert-parallel: true | ||
| enable-ep-weight-filter: true | ||
| moe-backend: deep_gemm_mega_moe | ||
| speculative-config: '{"method":"mtp","num_speculative_tokens":2}' | ||
| attention-config: '{"use_fp4_indexer_cache":true}' | ||
| max-model-len: 16384 | ||
| max-num-seqs: 512 | ||
| max-cudagraph-capture-size: 512 | ||
| max-num-batched-tokens: 1024 | ||
| trust-remote-code: true | ||
| no-enable-prefix-caching: true | ||
| no-enable-flashinfer-autotune: true | ||
| block-size: 256 | ||
| compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}' | ||
| gpu-memory-utilization: 0.9 | ||
| stream-interval: 50 | ||
| no-disable-hybrid-kv-cache-manager: true | ||
| enable-sleep-mode: true | ||
| tokenizer-mode: deepseek_v4 | ||
|
|
||
| benchmark: | ||
| type: "sa-bench" | ||
| isl: 8192 | ||
| osl: 1024 | ||
| concurrencies: "64x128x1024" | ||
| req_rate: "inf" | ||
| use_chat_template: true | ||
| custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer" | ||
|
|
||
116 changes: 116 additions & 0 deletions
116
...s/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-1p1d-dep4-tp4-mtp2.yaml
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 +1,116 @@ | ||
| name: "svf-vllm-disagg-gb300-1p1d-dep4-tp4-mtp2" | ||
|
|
||
| model: | ||
| path: "deepseek-v4-pro" | ||
| container: "vllm/vllm-openai:v0.21.0-ubuntu2404" | ||
| precision: "fp4" | ||
|
|
||
| dynamo: | ||
| install: true | ||
| wheel: "1.2.0.dev20260426" | ||
|
|
||
| setup_script: vllm-container-deps.sh | ||
|
|
||
| slurm: | ||
| time_limit: "8:00:00" | ||
|
|
||
| health_check: | ||
| max_attempts: 1440 | ||
| interval_seconds: 10 | ||
|
|
||
| resources: | ||
| gpu_type: "gb300" | ||
| gpus_per_node: 4 | ||
| prefill_nodes: 1 | ||
| decode_nodes: 1 | ||
| prefill_workers: 1 | ||
| decode_workers: 1 | ||
| gpus_per_prefill: 4 | ||
| gpus_per_decode: 4 | ||
|
|
||
| infra: | ||
| etcd_nats_dedicated_node: true | ||
|
|
||
| frontend: | ||
| type: dynamo | ||
| enable_multiple_frontends: false | ||
|
|
||
| backend: | ||
| type: vllm | ||
| connector: null | ||
| prefill_environment: | ||
| VLLM_ENGINE_READY_TIMEOUT_S: "3600" | ||
| TILELANG_CLEANUP_TEMP_FILES: "1" | ||
| VLLM_USE_NCCL_SYMM_MEM: "1" | ||
| NCCL_CUMEM_ENABLE: "1" | ||
| NCCL_MNNVL_ENABLE: "1" | ||
| NCCL_NVLS_ENABLE: "1" | ||
| TORCH_SYMMMEM: "NVSHMEM" | ||
| decode_environment: | ||
| VLLM_ENGINE_READY_TIMEOUT_S: "3600" | ||
| TILELANG_CLEANUP_TEMP_FILES: "1" | ||
| VLLM_USE_NCCL_SYMM_MEM: "1" | ||
| NCCL_CUMEM_ENABLE: "1" | ||
| NCCL_MNNVL_ENABLE: "1" | ||
| NCCL_NVLS_ENABLE: "1" | ||
| TORCH_SYMMMEM: "NVSHMEM" | ||
|
|
||
| vllm_config: | ||
| prefill: | ||
| kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' | ||
| served-model-name: "deepseek-ai/DeepSeek-V4-Pro" | ||
| kv-cache-dtype: "fp8" | ||
| tensor-parallel-size: 1 | ||
| pipeline-parallel-size: 1 | ||
| data-parallel-size: 4 | ||
| data-parallel-rpc-port: 13345 | ||
| enable-expert-parallel: true | ||
| enable-ep-weight-filter: true | ||
| moe-backend: deep_gemm_mega_moe | ||
| enforce-eager: true | ||
| attention-config: '{"use_fp4_indexer_cache":true}' | ||
| max-model-len: 16384 | ||
| max-num-seqs: 256 | ||
| max-num-batched-tokens: 16384 | ||
| trust-remote-code: true | ||
| no-enable-prefix-caching: true | ||
| no-enable-flashinfer-autotune: true | ||
| no-async-scheduling: true | ||
| block-size: 256 | ||
| gpu-memory-utilization: 0.9 | ||
| no-disable-hybrid-kv-cache-manager: true | ||
| enable-sleep-mode: true | ||
| tokenizer-mode: deepseek_v4 | ||
| decode: | ||
| kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' | ||
| served-model-name: "deepseek-ai/DeepSeek-V4-Pro" | ||
| kv-cache-dtype: "fp8" | ||
| tensor-parallel-size: 4 | ||
| pipeline-parallel-size: 1 | ||
| enable-expert-parallel: false | ||
| speculative-config: '{"method":"mtp","num_speculative_tokens":2}' | ||
| attention-config: '{"use_fp4_indexer_cache":true}' | ||
| max-model-len: 16384 | ||
| max-num-seqs: 512 | ||
| max-cudagraph-capture-size: 512 | ||
| max-num-batched-tokens: 1024 | ||
| trust-remote-code: true | ||
| no-enable-prefix-caching: true | ||
| no-enable-flashinfer-autotune: true | ||
| block-size: 256 | ||
| compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}' | ||
| gpu-memory-utilization: 0.9 | ||
| stream-interval: 50 | ||
| no-disable-hybrid-kv-cache-manager: true | ||
| enable-sleep-mode: true | ||
| tokenizer-mode: deepseek_v4 | ||
|
|
||
| benchmark: | ||
| type: "sa-bench" | ||
| isl: 8192 | ||
| osl: 1024 | ||
| concurrencies: "1x4x8" | ||
| req_rate: "inf" | ||
| use_chat_template: true | ||
| custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer" | ||
|
|
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.
🔴 All three new GB300 disagg MTP recipes set
speculative-config: '{"method":"mtp","num_speculative_tokens":2}'only on the decode worker, while every pre-existing DSV4 MTP disagg recipe in this directory (disagg-gb200-low-latency-mtp2.yaml, disagg-gb200-mid-curve-megamoe-mtp2.yaml, disagg-gb200-high-tpt-megamoe-mtp2.yaml) sets it on both prefill and decode. This looks like a copy-paste oversight where the option was added to decode but missed on the prefill side. Affectsdisagg-gb300-1p1d-dep4-tp4-mtp2.yaml(line 91),disagg-gb300-1p1d-dep4-dep8-megamoe-mtp2.yaml(line 95) anddisagg-gb300-2p1d-dep4-dep8-megamoe-mtp2.yaml(line 95) — fix by adding the samespeculative-configline to eachvllm_config.prefillblock (the GB200 recipes place it right afterenforce-eager: true).Extended reasoning...
What the bug is
The three newly added GB300 disagg MTP recipe files use
NixlConnectorwithkv_role: kv_bothto transfer the KV cache from prefill workers to decode workers, but only the decodevllm_configblock containsspeculative-config: '{"method":"mtp","num_speculative_tokens":2}'. The prefillvllm_configblock has no MTP setting at all, so the prefill worker boots a model graph without the MTP modules while the decode worker boots one with them.How the pattern was established
All three pre-existing DSV4 MTP disagg recipes in the same directory consistently set
speculative-configin both prefill and decode sections:disagg-gb200-low-latency-mtp2.yaml: lines 78 (prefill) and 99 (decode)disagg-gb200-mid-curve-megamoe-mtp2.yaml: lines 85 (prefill) and 112 (decode)disagg-gb200-high-tpt-megamoe-mtp2.yaml: lines 87 (prefill) and 113 (decode)The new GB300 files only set it in decode:
disagg-gb300-1p1d-dep4-tp4-mtp2.yaml: line 91 (decode only)disagg-gb300-1p1d-dep4-dep8-megamoe-mtp2.yaml: line 95 (decode only)disagg-gb300-2p1d-dep4-dep8-megamoe-mtp2.yaml: line 95 (decode only)The structural skeleton of the GB300 prefill blocks (same
enforce-eager: true,moe-backend: deep_gemm_mega_moe,enable-expert-parallel) is the same as the GB200 MTP recipes — they look exactly like a copy of the existing GB300 non-MTP recipe with MTP added only on the decode side.Why it matters
For DeepSeek MTP, the speculative module(s) are extra transformer blocks layered on top of the base model. Enabling MTP affects (a) which layers are loaded and (b) the KV-cache layout exposed to the connector. With
kv_role: kv_bothover NixlConnector the prefill side serializes KV state for the decode side to consume, so the two sides must agree on the model graph and its KV layout. If prefill runs without MTP while decode runs with MTP, the transferred KV blocks either cover fewer layers than decode expects or differ in shape — at best this breaks the benchmark, at worst it silently produces incorrect numbers from a file literally named*-mtp2.yaml. Even in cases where vLLM tolerates the divergence at runtime, this is a benchmark recipe and the resulting numbers would not represent the configuration the filename advertises.Proof / step-by-step
disagg-gb300-1p1d-dep4-dep8-megamoe-mtp2.yaml. Thebackend.vllm_config.prefillsection runs from lines 59–83;speculative-configis absent.backend.vllm_config.decodesection runs from lines 84–110; line 95 setsspeculative-config: '{"method":"mtp","num_speculative_tokens":2}'.kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'— i.e. prefill→decode KV transfer is active.disagg-gb200-high-tpt-megamoe-mtp2.yaml: same disagg-MTP shape, butspeculative-configis present on the prefill side (line 87) and decode side (line 113).disagg-gb300-1p1d-dep4-tp4-mtp2.yamland line 95 ofdisagg-gb300-2p1d-dep4-dep8-megamoe-mtp2.yaml. There is no obvious GB300-specific reason for the divergence; the difference is GPU type, not MTP semantics.How to fix
For each of the three new files, add the line
to the
backend.vllm_config.prefillblock (matching the convention in the GB200 MTP recipes, where it appears immediately afterenforce-eager: true). If the omission is intentional for some GB300-specific reason, a brief comment in the recipe explaining why would prevent future maintainers from "fixing" it.