Loading evidence and metric snapshots
Loading evidence and metric snapshots
rocm
Coverage, freshness, public chains, and sanitized evidence for sglang.
Live candidate for Scaled dot-product attention in sgl-project/sglang has 5 of 10 observed or reviewed stages and 1 of 9 verified edges. Missing relationships remain typed gaps.
InferenceX benchmark result 93094
The public InferenceX API reports 92.94321217652133 output tokens per second per GPU for DeepSeek R1 on eight MI355X GPUs at concurrency four, with 1,000 input and 1,000 output tokens.
sha256:09f1a7084285486fed350da4d22d7af2c4c5d853cd89a9a596a651187fce6888
InferenceX run 21407590933 attempt 2
The benchmark workflow completed successfully for InferenceX merge commit 2a796d10eca68071e5830c2d1a9a00287cbf51c0.
sha256:32ba269d2b318c51936ea8321621e94548d5b9f720bc4cf235421d7b01b6eaae
InferenceX PR 572 adds SGLang v0.5.8 MI355X results
The InferenceX changelog links SGLang PR 17327 and the dsr1-fp8-mi355x-sglang configuration.
sha256:dff011e3b2e496aa199ac9ef5b79b3eddca1bd16908af081dfd20aa31d86b258
InferenceX performance changelog linkage
The public changelog ties the benchmark update to SGLang PR 17327 and the v0.5.8 configuration.
sha256:b8cabcec722d790c4bf2dac711394c8aeaae733413ca348f97dee7b962bfe832
InferenceX SGLang MI355X benchmark configuration
The configuration pins image lmsysorg/sglang:v0.5.8-rocm700-mi35x for the DeepSeek R1 FP8 MI355X run.
sha256:eaeaf509683d968b1cb2feea3e70cfc6ec343af7ca5da6e4a5da021c17f5d91a
Release v0.5.8
Release observed
sha256:d2d1d4982279964c7f81754028bd57717ea6883ed4040a24d8abfe7d17db3612
SGLang v0.5.8
SGLang published v0.5.8 from commit 0189f41c30ede088a040a711a384f3024b8d7af5.
sha256:1cd9b9bdb52c4d29e3d53e78243284fc2301ff1654c0a56723d458a30e857ff1
SGLang v0.5.8 contains PR 17327 merge commit
GitHub compare reports the v0.5.8 commit is 73 commits ahead and zero commits behind merge commit 6988a0f5706c05ebc40ba9ef5c09243bcbf9886c.
sha256:0542e7a91617de186255aa2de36a51c0815d2b1835d62417738a88e4742786c3
SGLang PR 17327 disables the persistent MLA kernel without FP8 KV cache
Merged SGLang change 6988a0f5706c05ebc40ba9ef5c09243bcbf9886c updates the ROCm AITER attention backend guard.
sha256:95155a85c57c66b0935e33dd53407de965e70a4cbe822c0000dde3f3d75b9f01
SGLang AMD DeepSeek R1 MXFP4 eight-GPU test definition
The registered AMD test requires GSM8K accuracy above 0.94 and speed above 75 output tokens per second.
sha256:b6de7bfce3039aef3973cf12032f0cc705dba93782d10a5e8945273a8636effd
SGLang AMD pull request workflow
The PR workflow runs the registered AMD test suite in three automatically selected partitions.
sha256:b8b04df86423d0afae58c4281f831e473e3db4b363f410047d36382924bfd5c3
SGLang PR Test AMD run 21151393674 attempt 2
GitHub Actions reports attempt 2 completed successfully for PR head 35443039efe9c82aa73af6e8692fe7b2c5bd308d.
sha256:0d700cb8a0f3462740d0970475cc3ded8b82d685bfd6e6bb6ed234e685eaef14
SGLang AMD suite partitions for run 21151393674 attempt 2
GitHub reports success for AMD test partitions 0, 1, and 2. Retained artifacts are unavailable, so this evidence does not claim an exact file partition or assertion value.
sha256:e1ca050740f74d976c6c0b0f231721d93459a199e1d416384ead82e8b0b1c1b9