[NVIDIA] update NVIDIA single-node DeepSeek R1 SGLang image from 0.5.8 to latest 0.5.9

[cquil11]

For each SKU,

• Search SGLang release notes / commit history to check what changes must be made to recipes since the 0.5.8 release (for instance, https://github.com/sgl-project/sglang/releases/tag/v0.5.9)
• Search SGLang cookbook recipes to see if any updates have been made. Although this is a good reference, these cookbooks are often outdated and it is better to rely on the actual release commit history.
• Perform e2e test confirming the image works and perf is stable
• Update perf-changelog.yaml accordingly
• Only update single-node configs

[cquil11]

@claude Please lay the groundwork for this issue and complete steps laid out

[Klaud-Cold]

Claude finished @cquil11's task in 58m 50s —— View job • claude/issue-813-20260226-1557 • [Create PR ➔](https://github.com/SemiAnalysisAI/InferenceX/compare/main...claude/issue-813-20260226-1557?quick_pull=1&title=Issue%20%23813%3A%20Changes%20from%20Claude&body=This%20PR%20addresses%20issue%20%23813%0A%0AGenerated%20with%20%5BClaude%20Code%5D(https%3A%2F%2Fclaude.ai%2Fcode)

---

Updating NVIDIA Single-Node DeepSeek R1 SGLang Image to 0.5.9

• Gather repo context and identify configs
• Research SGLang 0.5.9 release changes
• Verify Docker image tags for v0.5.9
• Update image tags in nvidia-master.yaml
• Fix deprecated SGL_ env vars in benchmark scripts
• Update perf-changelog.yaml
• Commit and push changes
• Run e2e tests to confirm image works and perf is stable (queued, waiting for runners)
• Analyze benchmark results and finalize

---

Changes made (commit 58e333c):

Config	Previous Image	New Image
dsr1-fp4-b200-sglang	v0.5.6-cu129-amd64	v0.5.9-cu129-amd64
dsr1-fp8-b200-sglang	v0.5.6-cu129-amd64	v0.5.9-cu129-amd64
dsr1-fp8-b200-sglang-mtp	v0.5.8-cu130-amd64	v0.5.9-cu130-amd64
dsr1-fp8-h200-sglang	v0.5.7-cu129-amd64	v0.5.9-cu129-amd64

Additional fixes:

• Fixed deprecated SGL_ENABLE_JIT_DEEPGEMM → SGLANG_ENABLE_JIT_DEEPGEMM in benchmarks/single_node/dsr1_fp8_b200.sh

Key SGLang 0.5.9 improvements:

• TRT-LLM NSA kernel integration (3-5x speedup for DeepSeek on Blackwell)
• Flashinfer upgraded to 0.6.3
• SpecV2 out-of-index bug fix (torch GC)
• MoE fused kernel optimizations, radix cache eviction improvements

E2E test run 1: Run #22450221051 — cancelled

• fp4 tp=4 ep=4 job was working (server started successfully, serving requests) before external cancellation
• 2 jobs failed with no logs (b200-dgxc-slurm runner infra issue, not image-related)

E2E test run 2: Run #22451520861 — queued (waiting for GPU runners)

Summary: All code changes are complete and pushed. PR #814 is open. The first test run confirmed the v0.5.9 image loads and starts serving correctly (fp4 b200 job was operational before cancellation). The retry run is queued and awaiting available GPU runners to complete the full benchmark validation. Results should be reviewed once runners become available.
|