[Leonccaa]

Change

Updates the README and guides for 0.1.4 and adds the measurements taken after the release. No code changes.

• README: 0.1.4 is the current release (install commands and image tag). The capacity table uses the recommended probabilistic DFlash2 settings: 1.37 million tokens (2.61× at 512K); the 1.38 million / 2.63× it showed before is the greedy-drafting value. A cached 32K-token document now comes back from CPU memory in 0.6 s, and from a local NVMe disk tier after an engine restart in 1.5 s, against 10.1 s recomputed. The experimental whole-table PLE disk option is described with its measured cost. "Next" lists the 0.1.5 plans. Thanks to the authors of NVLink results (V100-SXM2-32GB NVLink quad, TP4+DCP2, DFlash2) for 0.1.3, from a 9-model serving bench #38 and Nvfp4 Test Results #37.
• Positioning: ReVolta is built for 4 × V100 32 GB; NVLink and one PCIe switch with P2P are the recommended and tested interconnects (README, docs/known-limits.md, .github/SUPPORT.md and the bug report form).
• docs/known-limits.md: the whole-table PLE disk option measured on 0.1.4 on local NVMe and SATA SSDs, under 64 and 32 GiB limits, with and without a 32 GiB CPU KV pool: single-request decoding about 23–25% slower, a cold first token 1.7–2.2 times as long, a warm first token within 2–3% of the pinned path (5–8% with the CPU KV pool), and greedy output identical to the pinned path.
• docs/kv-offload.md: with DFlash2, the drafter's cache block size differs from the target's, so the CPU pool and the disk tier give every cache group pages of the full block size: about 50 KB per token, about 2.5 times MTP4. The sizing rule for disk restores now gives both values. Adds the local NVMe results after a restart.
• docs/performance.md: a 0.1.4 readback table (recomputed, from CPU memory, from NVMe after a restart) at 32K, 99K and 131K tokens; the earlier measurement stays below it.

Support matrix: NVLink changes from "community results, outside our release gate" to "recommended; tested with community results" in the README and .github/SUPPORT.md. Nothing else in the matrix changes.

Validation

• Readback numbers: 4 × V100 32 GB PCIe, 0.1.4 with DFlash2 (recommended settings), Uncensored AWQ, E4M3 KV, 250 W, pinned PLE table, 7 GiB CPU pool and a 64 GiB disk budget on a local NVMe SSD (ext4). The page cache was verified empty before the restart (1,088 files, 27.8 GB, 0 resident). 11 of 11 documents were read back with the whole cacheable prefix and the same 64-token greedy output as the cold run.
• DFlash2 page size: vllm/v1/kv_offload/cpu/spec.py uses the non-partitioned layout when the cacheable groups have different GPU block sizes. The disk pages written per document match it: 72 pages of 25.6 MB (1.84 GB) for 32K tokens, 249 pages (6.36 GB) for 131K.
• PLE disk numbers: six arms in one window on 0.1.4 (pinned baseline; NVMe and SATA under 64 GiB; NVMe under 32 GiB; NVMe and SATA under 64 GiB with a 32 GiB CPU KV pool), 250 W, three 64-token greedy prompts identical to the baseline in every arm.
• Capacity: 0.1.4 release gate logs (GPU KV cache size: 1,366,121 tokens, 2.61x with probabilistic DFlash2; MTP4 unchanged at 1,153,143).
• pre-commit run --files on the six changed files (pre-commit 4.7.0): all hooks pass, including typos and markdownlint-cli2. The bug report form parses as YAML.

Attribution

AI-assisted: drafted with Claude Code. The commit is signed off under the Developer Certificate of Origin. A human review of every changed line is pending before merge.

🤖 Generated with Claude Code