K3 blog update

[xianbaoqian]

No description provided.

[cloudflare-workers-and-pages]

Deploying vllm-blog-source with    Cloudflare Pages

Latest commit:	ec0f1b5
Status:	✅  Deploy successful!
Preview URL:	https://c47a96b3.vllm-blog-source.pages.dev
Branch Preview URL:	https://k32.vllm-blog-source.pages.dev

View logs

[chatgpt-codex-connector]

💡 Codex Review

vllm-project.github.io/_posts/2026-07-27-k3.md

Lines 271 to 272 in 87ffaae

	--nnodes 2 \
	--node-rank 0 \

Include the missing worker launch for the two-node server

This reproduction block declares --nnodes 2 but only shows the rank-0 server and leaves --master-addr at its localhost default. In vLLM's multiprocessing multi-node launch, the other node must also be started with its own --node-rank 1 --headless --master-addr <HEAD_NODE_IP> command; copied as-is, the head rank waits for a missing peer before any benchmark can run. Please either include the worker-node command or make this a single-node TP8 recipe.

---

vllm-project.github.io/_posts/2026-07-27-k3.md

Line 282 in 87ffaae

--speculative-config '{"model":"Inferact/Kimi-K3-DSpark","method":"dspark","num_speculative_tokens":7,"attention_backend":"FLASHINFER_MLA","draft_sample_method":"probabilistic","rejection_sample_method":"block"}'

Split the benchmark server runs by speculative mode

Because this vllm serve command is launched with --speculative-config, the immediately following vllm-bench run labeled “no speculative decoding” still measures the DSpark/speculative path. That makes the reproduction instructions unable to reproduce the non-DSpark throughput numbers; please show a separate serve command without this flag for the no-spec run, or move this flag to a second server run used only for the DSpark benchmark.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".