Repository navigation
[K3.1-Pro] Day-0 support vLLM on MI355X / [K3.1-Pro] MI355X 上 vLLM 的 Day-0 支持 #4806
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
Merged
Merged
Changes from all commits
Commits
Show all changes
32 commits
Select commit
Hold shift + click to select a range
06e3ea7
feat(kimik3.1): add Kimi K3.1 Pro vLLM AgentX on MI355X
cquil11 de1f23a
chore(kimik3.1): link the MI355X vLLM changelog entry to its PR
cquil11 bc5c564
fix(kimik3.1): keep only documented or common recipe settings
cquil11 6a8fc96
Merge remote-tracking branch 'origin/main' into cam/kimik3.1-mi355x-vllm
cquil11 bda25c5
fix(kimik3.1): leave GPU headroom for pinned scratch on MI355X
cquil11 aa2bd20
perf(kimik3.1): match the apex-1104 sweep's points and batching
billishyahao 4594456
Merge remote-tracking branch 'origin/main' into cam/kimik3.1-mi355x-vllm
cquil11 de31629
fix(kimik3.1): resolve the MI355X DRAM budget to the recipe's 2577 GB
cquil11 2821c4b
perf(kimik3.1): use 7 DSpark tokens and the specialized decode at conc 1
billishyahao 24f26ac
perf(kimik3.1): prefill with the AITER FA MLA backend at conc 48 only
billishyahao ba19e2a
perf(kimik3.1): set breakable cudagraphs and scratch reclaim per point
billishyahao f7e7056
Merge remote-tracking branch 'origin/main' into cam/kimik3.1-mi355x-vllm
cquil11 800b3ca
Merge remote-tracking branch 'origin/main' into cam/kimik3.1-mi355x-vllm
cquil11 d3f324b
fix(kimik3.1): keep scratch reclaim off on the MI355X DRAM points
cquil11 3e1771f
perf(kimik3.1): trim the conc-32 and conc-48 points on MI355X
billishyahao 44c11f8
perf(kimik3.1): keep scratch allocated and step the draft count down
billishyahao fcbdb39
Merge remote-tracking branch 'inf/main' into cam/kimik3.1-mi355x-vllm
billishyahao 641e8f5
Merge branch 'main' into cam/kimik3.1-mi355x-vllm
billishyahao 9d80dc7
Merge branch 'main' into cam/kimik3.1-mi355x-vllm
billishyahao d843db2
Merge remote-tracking branch 'origin/main' into cam/kimik3.1-mi355x-vllm
cquil11 a164a99
fix(synthetic-acceptance): drop Kimi vLLM parsers on synthetic throug…
cquil11 db2754b
Merge remote-tracking branch 'origin/main' into cam/kimik3.1-mi355x-vllm
cquil11 18303be
fix(kimik3.1-mi355x-vllm): lower gpu-memory-utilization to 0.86 at co…
cquil11 33d17a1
Merge remote-tracking branch 'origin/main' into cam/kimik3.1-mi355x-vllm
cquil11 7596fb6
fix
billishyahao 6cdc164
Merge branch 'main' into cam/kimik3.1-mi355x-vllm
billishyahao 14a3dc8
fix
billishyahao e1beac3
perf(kimik3.1): refresh the day-0 image and map the offload pool with…
billishyahao 740ee08
Merge branch 'main' into cam/kimik3.1-mi355x-vllm
billishyahao cf9f666
docs(changelog): sync Kimi-K3.1-Pro MI355X vLLM block with the diff […
seungrokj 312a967
Merge branch 'main' into cam/kimik3.1-mi355x-vllm [skip-sweep]
seungrokj b1c4fce
chore: refresh PR #4806 for sweep reuse [skip-sweep]
cquil11 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
198 changes: 198 additions & 0 deletions
198
...ex-e2e/benchmarks/single_node/srt-slurm-recipes/kimik3.1/vllm/mi355x-fp4-mtp/agentic.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 +1,198 @@ | ||
| base: | ||
| schema: 2 | ||
| name: kimik3.1-fp4-mi355x-vllm-agentic | ||
| resources: | ||
| gpu_type: mi355x | ||
| gpus_per_node: 8 | ||
| frontend: | ||
| type: vllm | ||
| enable_multiple_frontends: false | ||
| observability: | ||
| enabled: false | ||
| tachometer: | ||
| enabled: false | ||
| engine: | ||
| type: vllm | ||
| connector: null | ||
| # A node's first run downloads ~1.56 TB of MXFP4 shards into its local HF cache before loading. | ||
| health_check: | ||
| interval_seconds: 10 | ||
| max_attempts: 720 | ||
| roles: | ||
| agg: | ||
| nodes: 1 | ||
| workers: 1 | ||
| gpus: 8 | ||
| args: | ||
| served-model-name: moonshotai/Kimi-K3.1-Pro | ||
| tensor-parallel-size: 8 | ||
| decode-context-parallel-size: 8 | ||
| dcp-comm-backend: a2a | ||
| cp-kv-cache-interleave-size: 1 | ||
| trust-remote-code: true | ||
| moe-backend: auto | ||
| seed: 7 | ||
| language-model-only: true | ||
| load-format: fastsafetensors | ||
| enable-auto-tool-choice: true | ||
| tool-call-parser: kimi_k3 | ||
| reasoning-parser: kimi_k3 | ||
| max-model-len: 1048576 | ||
| enable-prefix-caching: true | ||
| prefix-match-unit: 64 | ||
| kv-cache-dtype: fp8 | ||
| # The K3.1 week-1 precision policy requires an FP32 KDA recurrent state. | ||
| mamba-ssm-cache-dtype: float32 | ||
| # The policy's FP4 indexer cache; the image's AITER pads the FP4 indexer pages it needs. | ||
| attention-config: '{"indexer_kv_dtype":"mxfp4"}' | ||
| additional-config: '{"kda_prefill_backend":"fused"}' | ||
| # Prefix-cached prefill chunks end on K3.1's 1152-token packed DSA block, so the budget floors to a multiple of 1152. | ||
| max-num-batched-tokens: 16384 | ||
| # Kernel scratch, up to HSA_SCRATCH_SINGLE_LIMIT_ASYNC per dispatch, is allocated outside this budget, so leave headroom. | ||
| gpu-memory-utilization: 0.90 | ||
| stream-interval: 10 | ||
| disable-uvicorn-access-log: true | ||
| # 224.875 GB per rank; TOTAL_CPU_DRAM_GB (1799) split across the 8 TP ranks. | ||
| kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use_per_rank":224875000000,"lazy_offload":false}}' | ||
| env: | ||
| VLLM_ROCM_USE_AITER: '1' | ||
| VLLM_ROCM_USE_AITER_MOE_SITUV2: '1' | ||
| SAFETENSORS_FAST_GPU: '1' | ||
| VLLM_USE_V2_MODEL_RUNNER: '1' | ||
| VLLM_FASTSAFETENSORS_QUEUE_SIZE: '-1' | ||
| # K3.1 has no torch.compile submodule, so only breakable graphs capture the prefill | ||
| # sizes the throughput points list. The conc-1 override turns this off. | ||
| VLLM_USE_BREAKABLE_CUDAGRAPH: '1' | ||
| VLLM_ROCM_AITER_MLA_ASM_PADDING: asm | ||
| # Verify the draft tokens with the AITER assembly DCP kernel. | ||
| VLLM_ROCM_AITER_MLA_DCP_VERIFY: asm | ||
| VLLM_ROCM_DCP_AITER_ALLGATHER: '1' | ||
| VLLM_ROCM_USE_AITER_FLYDSL_SPARSE_MLA: '1' | ||
| # AITER reads its own variable for the same kernels. | ||
| AITER_FLYDSL_SPARSE_MLA: '1' | ||
| VLLM_ROCM_DSA_DCP_GLOBAL_KEY_PREFILL: '1' | ||
| VLLM_ROCM_DSA_DCP_ROWSHARD_PREFILL: '1' | ||
| VLLM_USE_DIRECT_DCP_A2A: '0' | ||
| VLLM_USE_DIRECT_DCP_Q_GATHER: '0' | ||
| VLLM_USE_DIRECT_DCP_KV_GATHER: '0' | ||
| # The K3.1 policy and MODELS.md keep all-reduce in BF16, so quick reduce must not quantize. | ||
| VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: NONE | ||
| AITER_FLYDSL_STAGE2_FP8: '0' | ||
| AITER_BF16_FP8_MOE_BOUND: '0' | ||
| HSA_NO_SCRATCH_RECLAIM: '1' | ||
| # Back the offload pool with transparent huge pages. Torch's CPU allocator | ||
| # only calls madvise(MADV_HUGEPAGE) when this is set, and the host is in | ||
| # madvise mode, so without it the pool is all 4 KiB pages. | ||
| THP_MEM_ALLOC_ENABLE: '1' | ||
| VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: '1200' | ||
| PYTHONNOUSERSITE: '1' | ||
| VLLM_ENGINE_READY_TIMEOUT_S: '7200' | ||
| # Identical prefixes must hash to identical block keys run-to-run. | ||
| PYTHONHASHSEED: '42' | ||
| benchmark: | ||
| env: | ||
| AIPERF_HTTP_TCP_USER_TIMEOUT: '900000' | ||
| AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'vllm:' | ||
|
|
||
| # max-num-seqs stays below CONC from c4 to c32, and equals it at c1 and c48. The capture list is | ||
| # the decode sizes, multiples of one plus the draft token count, then the prefill ladder. | ||
| # Throughput runs replace block rejection with the golden acceptance length. | ||
| override_gpu_kv_c1: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-num-seqs: 1 | ||
| compilation-config: '{"mode":3,"cudagraph_mode":"FULL_AND_PIECEWISE","max_cudagraph_capture_size":16,"custom_ops":["+fused_rms_norm_gated"],"cudagraph_capture_sizes":[2,4,6,8,10,12,14,16]}' | ||
| additional-config: '{"decode_backend":"latency_specialized","kda_prefill_backend":"fused"}' | ||
| speculative-config: '{"method":"dspark","model":"Inferact/Kimi-K3.1-Pro-DSpark-preview","num_speculative_tokens":7,"attention_backend":"ROCM_AITER_MLA","kv_cache_dtype":"fp8","draft_sample_method":"probabilistic","rejection_sample_method":"block"}' | ||
| kv-transfer-config: null | ||
| env: | ||
| # This point captures only the decode sizes, so breakable graphs buy nothing. | ||
| VLLM_USE_BREAKABLE_CUDAGRAPH: '0' | ||
| benchmark: | ||
| env: | ||
| CONC: '1' | ||
| KV_OFFLOADING: 'none' | ||
|
|
||
| override_dram_c4: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-num-seqs: 2 | ||
| compilation-config: '{"mode":3,"cudagraph_mode":"FULL_AND_PIECEWISE","max_cudagraph_capture_size":16384,"custom_ops":["+fused_rms_norm_gated"],"cudagraph_capture_sizes":[6,12,16,24,32,48,64,96,128,160,192,224,256,288,320,352,384,416,448,480,512,640,768,896,1024,1280,1536,1792,2048,2560,3072,3584,4096,4608,5120,6144,6400,7168,8192,8704,10240,12288,14336,16384]}' | ||
| speculative-config: '{"method":"dspark","model":"Inferact/Kimi-K3.1-Pro-DSpark-preview","num_speculative_tokens":5,"attention_backend":"ROCM_AITER_MLA","kv_cache_dtype":"fp8","draft_sample_method":"probabilistic","rejection_sample_method":"block"}' | ||
| benchmark: | ||
| env: | ||
| CONC: '4' | ||
| KV_OFFLOADING: 'dram' | ||
| TOTAL_CPU_DRAM_GB: '1799' | ||
|
|
||
| override_dram_c10: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-num-seqs: 6 | ||
| compilation-config: '{"mode":3,"cudagraph_mode":"FULL_AND_PIECEWISE","max_cudagraph_capture_size":16384,"custom_ops":["+fused_rms_norm_gated"],"cudagraph_capture_sizes":[6,12,18,24,30,36,48,64,96,128,160,192,224,256,288,320,352,384,416,448,480,512,640,768,896,1024,1280,1536,1792,2048,2560,3072,3584,4096,4608,5120,6144,6400,7168,8192,8704,10240,12288,14336,16384]}' | ||
| speculative-config: '{"method":"dspark","model":"Inferact/Kimi-K3.1-Pro-DSpark-preview","num_speculative_tokens":5,"attention_backend":"ROCM_AITER_MLA","kv_cache_dtype":"fp8","draft_sample_method":"probabilistic","rejection_sample_method":"block"}' | ||
| benchmark: | ||
| env: | ||
| CONC: '10' | ||
| KV_OFFLOADING: 'dram' | ||
| TOTAL_CPU_DRAM_GB: '1799' | ||
|
|
||
| override_dram_c16: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-num-seqs: 10 | ||
| compilation-config: '{"mode":3,"cudagraph_mode":"FULL_AND_PIECEWISE","max_cudagraph_capture_size":16384,"custom_ops":["+fused_rms_norm_gated"],"cudagraph_capture_sizes":[5,10,15,20,25,30,35,40,45,50,64,96,128,160,192,224,256,288,320,352,384,416,448,480,512,640,768,896,1024,1280,1536,1792,2048,2560,3072,3584,4096,4608,5120,6144,6400,7168,8192,8704,10240,12288,14336,16384]}' | ||
| speculative-config: '{"method":"dspark","model":"Inferact/Kimi-K3.1-Pro-DSpark-preview","num_speculative_tokens":4,"attention_backend":"ROCM_AITER_MLA","kv_cache_dtype":"fp8","draft_sample_method":"probabilistic","rejection_sample_method":"block"}' | ||
| benchmark: | ||
| env: | ||
| CONC: '16' | ||
| KV_OFFLOADING: 'dram' | ||
| TOTAL_CPU_DRAM_GB: '1799' | ||
|
|
||
| override_dram_c24: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-num-seqs: 16 | ||
| compilation-config: '{"mode":3,"cudagraph_mode":"FULL_AND_PIECEWISE","max_cudagraph_capture_size":16384,"custom_ops":["+fused_rms_norm_gated"],"cudagraph_capture_sizes":[5,10,15,20,25,30,35,40,45,50,55,60,65,70,75,80,96,128,160,192,224,256,288,320,352,384,416,448,480,512,640,768,896,1024,1280,1536,1792,2048,2560,3072,3584,4096,4608,5120,6144,6400,7168,8192,8704,10240,12288,14336,16384]}' | ||
| speculative-config: '{"method":"dspark","model":"Inferact/Kimi-K3.1-Pro-DSpark-preview","num_speculative_tokens":4,"attention_backend":"ROCM_AITER_MLA","kv_cache_dtype":"fp8","draft_sample_method":"probabilistic","rejection_sample_method":"block"}' | ||
| benchmark: | ||
| env: | ||
| CONC: '24' | ||
| KV_OFFLOADING: 'dram' | ||
| TOTAL_CPU_DRAM_GB: '1799' | ||
|
|
||
| override_dram_c32: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-num-seqs: 24 | ||
| max-num-batched-tokens: 10240 | ||
| gpu-memory-utilization: 0.91 | ||
| compilation-config: '{"mode":3,"cudagraph_mode":"FULL_AND_PIECEWISE","max_cudagraph_capture_size":10240,"custom_ops":["+fused_rms_norm_gated"],"cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64,68,72,76,80,84,88,92,96,128,160,192,224,256,288,320,352,384,416,448,480,512,640,768,896,1024,1280,1536,1792,2048,2560,3072,3584,4096,4608,5120,6144,6400,7168,8192,8704,10240]}' | ||
| speculative-config: '{"method":"dspark","model":"Inferact/Kimi-K3.1-Pro-DSpark-preview","num_speculative_tokens":3,"attention_backend":"ROCM_AITER_MLA","kv_cache_dtype":"fp8","draft_sample_method":"probabilistic","rejection_sample_method":"block"}' | ||
| benchmark: | ||
| env: | ||
| CONC: '32' | ||
| KV_OFFLOADING: 'dram' | ||
| TOTAL_CPU_DRAM_GB: '1799' | ||
|
|
||
| override_dram_c48: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-num-seqs: 48 | ||
| max-num-batched-tokens: 10240 | ||
| gpu-memory-utilization: 0.91 | ||
| compilation-config: '{"mode":3,"cudagraph_mode":"FULL_AND_PIECEWISE","max_cudagraph_capture_size":10240,"custom_ops":["+fused_rms_norm_gated"],"cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64,68,72,76,80,84,88,92,96,100,104,108,112,116,120,124,128,132,136,140,144,148,152,156,160,164,168,172,176,180,184,188,192,224,256,288,320,352,384,416,448,480,512,640,768,896,1024,1280,1536,1792,2048,2560,3072,3584,4096,4608,5120,6144,6400,7168,8192,8704,10240]}' | ||
| speculative-config: '{"method":"dspark","model":"Inferact/Kimi-K3.1-Pro-DSpark-preview","num_speculative_tokens":3,"attention_backend":"ROCM_AITER_MLA","kv_cache_dtype":"fp8","draft_sample_method":"probabilistic","rejection_sample_method":"block"}' | ||
| benchmark: | ||
| env: | ||
| CONC: '48' | ||
| KV_OFFLOADING: 'dram' | ||
| TOTAL_CPU_DRAM_GB: '1799' | ||
|
|
||
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.
🔴 On every MI355X node without the ~1.56 TB checkpoint already cached, the first K3.1 point can fail its readiness check. The download, load and graph capture all fit in the 7200 s budget only if the node downloads at about 217 MB/s or faster. Nothing routes this model to the shared cache. The budget is
max_attempts: 720x 10 s, matchingVLLM_ENGINE_READY_TIMEOUT_S: '7200'. Fix: put the checkpoint where each node does not re-download it before serving. For example, addmoonshotai/Kimi-K3.1-*toSHARED_HF_CACHE_LANES["mi355x-amds"], or stage amodels.entriesrow. Alternatively, size the readiness budget for a 1.56 TB cold download.Why this was flagged
The trigger is a sweep point landing on an mi355x-amds node whose node-local
/var/lib/hf-hub-cache(runners.yaml:789) does not yet hold Kimi-K3.1-Pro. Every node starts that way, because nomodels.entriesrow exists for this model on mi355x-amds (runners.yaml:776-780).single_node_hf_cache(infx/launch/drivers/srt/models.py:198-201) only picksshared-hf-hub-cachefor models matched inSHARED_HF_CACHE_LANES, and models.py:187-194 has no Kimi rule. As a result, vLLM pulls about 1.56 TB from HF inside server startup. The sibling MI355X recipe uses the same 720 x 10 s health check and notes at minimaxm31flash/vllm/mi355x-fp4/agentic.yaml:26-27 that a cold load of only 240 GB can use up the whole VLLM_ENGINE_READY_TIMEOUT_S. On a slow link, the first point on each node exceeds 7200 s and the job fails. Underfull-sweep-fail-fast, that one failure cancels the sweep. The 10 points spread across nodes each pay the download separately. The recipe's comment at line 17 acknowledges the download, but nothing bounds or avoids it.Verification: The readiness budget is 7200 s: agentic.yaml:18-20 sets
interval_seconds: 10andmax_attempts: 720. No shared-cache rule matchesmoonshotai/Kimi-K3.1-Pro(models.py:187-194), so it falls through to node-local/var/lib/hf-hub-cache(runners.yaml:789), and vLLM downloads it during startup. 1.56 TB / 7200 s ≈ 217 MB/s. A node below that rate fails its first point.