Leonccaa · Oct 26, 2026 · Oct 24, 2026 · Oct 24, 2026 · Oct 24, 2026 · Oct 24, 2026
diff --git a/.github/MULTIMODAL_INPUTS.md b/.github/MULTIMODAL_INPUTS.md
@@ -39,7 +39,7 @@ This V100 behaviour is the same as in 0.1.0 and 0.1.1, and 0.1.2 keeps it. Set
 
 - the generic 0.1.2 default: `--limit-mm-per-prompt '{"image": 8, "video": 1, "audio": 1}'`;
 - the reference deployment, as in the README launch command:
-  `--limit-mm-per-prompt '{"image": 16, "video": 1, "audio": 16}'`.
+  `--limit-mm-per-prompt '{"image": 16, "video": 1}'` (Qwen3.8 Flash-Next has no audio input).
 
 The image resizing below applies on V100 too.
 

diff --git a/.github/SUPPORT.md b/.github/SUPPORT.md
@@ -1,4 +1,4 @@
-# ReVolta 0.1.0 support scope
+# ReVolta support scope
 
 The [support matrix in the README](../README.md#support-matrix) is the reference.
 In short:
@@ -8,14 +8,14 @@ In short:
 | GPUs | 4 × NVIDIA V100 32 GB PCIe |
 | PCIe switch with P2P | Validated |
 | PCIe without P2P | Validated by disabling P2P on P2P-capable hardware; about 30–55% of P2P speed |
-| NVLink | Unverified (no hardware available); the upstream topology path and `VLLM_SM70_INTERCONNECT` override exist |
+| NVLink | Community results on 4 × V100-SXM2 32 GB ([#2](https://github.com/Leonccaa/ReVolta/issues/2)); outside our release gate |
 | Host memory | 128 GiB recommended; 80 GiB measured minimum with KV offload off |
 | KV offload | Off, CPU, or CPU + SSD tier |
 | KV cache | E4M3 (default) or INT8; INT8 with the SSD tier is not supported |
 | Models | Qwen3.8 Flash-Next (QSA, multimodal) |
 | Checkpoints | AWQ g32, standard NVFP4 `RadixArk/Qwen3.8-Flash-Next-NVFP4`, `nvidia/Qwen3.8-Flash-Next-FP4-Local` |
 
-Unverified configurations (NVLink, TP2/TP8, V100 16 GB, other models or
+Configurations outside our release gate (NVLink, TP2/TP8, V100 16 GB, other models or
 topologies) are handled on a best-effort basis. Other GPU architectures,
 ROCm/CPU/XPU/TPU serving and GGUF are not built.
 

diff --git a/BUILDING.md b/BUILDING.md
@@ -5,7 +5,7 @@ The release version is `RELEASE_VERSION` in `setup.py` (initially `0.1.0`).
 Set `VLLM_VERSION_OVERRIDE` for an exact wheel and runtime version, or pass
 `SETUPTOOLS_SCM_PRETEND_VERSION` to the Docker build. Inherited Git tags do not
 select the ReVolta version. Use a PEP 440 local version for private deployments,
-for example `0.1.0+deployment.20261018.g0123456789`.
+for example `0.1.3+deployment.20261025.g0123456789`.
 
 ## SM70 source build
 
@@ -15,13 +15,13 @@ Use CUDA 12.8; CUDA 13 cannot compile SM70. Limit parallelism to available memor
 NVCC thread per job are a conservative starting point on a 32 GB host.
 
 ```bash
-version=0.1.0
+version=0.1.3
 docker build --target build -f docker/Dockerfile \
   --build-arg torch_cuda_arch_list=7.0 \
   --build-arg max_jobs=4 --build-arg nvcc_threads=1 \
   --build-arg SETUPTOOLS_SCM_PRETEND_VERSION="$version" \
-  -t local/revolta-build:0.1.0 .
-container=$(docker create local/revolta-build:0.1.0)
+  -t "local/revolta-build:$version" .
+container=$(docker create "local/revolta-build:$version")
 mkdir -p artifacts
 docker cp "$container:/workspace/dist" artifacts/
 docker cp "$container:/workspace/debug" artifacts/
@@ -31,7 +31,7 @@ docker build --target vllm-openai -f docker/Dockerfile \
   --build-arg torch_cuda_arch_list=7.0 \
   --build-arg max_jobs=4 --build-arg nvcc_threads=1 \
   --build-arg SETUPTOOLS_SCM_PRETEND_VERSION="$version" \
-  -t local/revolta:0.1.0 .
+  -t "local/revolta:$version" .
 ```
 
 Flash-V100 and FlashQLA are bundled into the ReVolta wheel. DeepEP is skipped

diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
@@ -1,7 +1,7 @@
 # Contributing to ReVolta
 
 ReVolta focuses on the configurations in the
-[0.1.0 support scope](.github/SUPPORT.md), derived from roadmap W10. Inherited
+[support scope](.github/SUPPORT.md). Inherited
 vLLM functionality outside that scope is not a support commitment.
 
 ## Issues

diff --git a/README.md b/README.md
diff --git a/SECURITY.md b/SECURITY.md
@@ -3,7 +3,7 @@
 Report suspected vulnerabilities through this repository's GitHub
 **Security → Advisories → Report a vulnerability** form (private vulnerability
 reporting). Do not put exploit details, credentials or private data in a public
-issue. Private reporting must be enabled before the public release.
+issue.
 
 Include the affected version or commit, configuration, impact, reproduction
 steps and any suggested mitigation. Maintainers will coordinate investigation,

diff --git a/docs/checkpoints.md b/docs/checkpoints.md
diff --git a/docs/configuration.md b/docs/configuration.md
@@ -0,0 +1,127 @@
+# Configuration reference
+
+This page explains the reference deployment settings, sampling defaults and resource budgets.
+Use the [getting-started tutorial](getting-started.md) for the installation and launch sequence.
+
+## Environment settings
+
+The engine detects SM70, TP4, a QSA model (Qwen3.8 Flash-Next) and FP16, and then sets
+the tuned kernel switches by itself (the "ReVolta profile"). Only these deployment
+settings remain:
+
+The complete reference exports are in [step 3 of the tutorial](getting-started.md#3-environment).
+
+Use only the documented `VLLM_SM70_*` settings unless you are troubleshooting. Explicit
+values override the profile.
+
+## Interconnect
+
+The engine selects the interconnect automatically and logs `SM70 interconnect=<mode>`.
+P2P uses custom all-reduce for messages up to 80 KiB inside CUDA graphs; other
+collectives use NCCL. Override with `VLLM_SM70_INTERCONNECT=auto|pex|nccl|nvlink`
+or disable custom all-reduce with `--disable-custom-all-reduce`.
+
+## Launch options
+
+- **`draft_sample_method` must be `greedy`.** A probabilistic draft corrupts output
+  above `max_num_seqs=2` at temperature > 0 (upstream 1Cat-vLLM issue #656). This
+  setting is about draft verification only; clients can still use any sampling
+  parameters.
+- **`--gpu-memory-utilization 0.95`** passed the full gate, including the multimodal
+  stress test with 8 concurrent requests. Higher values ran out of memory in earlier
+  multimodal stress tests. If you change concurrency or media limits, measure again.
+- The scheduler settings (`--prefill-pressure-*`, `--long-prefill-token-threshold`,
+  `--mixed-prefill-step-latency-ms`) and `ple_disk_cascade: false` are the tested
+  values for this machine. They are deployment settings, not universal defaults.
+- `cpu_bytes_to_use` is the total for all four workers (48 GiB = 12 GiB per worker).
+- **Keep `--limit-mm-per-prompt` on V100.** With FlashAttention-V100, the engine allows
+  1 image and no video per request without it, and a value that leaves out `"video"`
+  disables video. 0.1.3 keeps this V100 default. `{"image": 16, "video": 1}`
+  keeps the reference deployment's limits; Qwen3.8 Flash-Next takes images and video,
+  not audio, so no audio limit is set. To use the generic image count instead, set
+  `--limit-mm-per-prompt '{"image": 8, "video": 1}'`; a lower image count leaves more
+  host memory for bursts of large images. Large images are resized
+  either way (see [Support and known limits](known-limits.md)).
+- `--structured-outputs-config '{"backend": "xgrammar"}'` pins the structured-output
+  backend. With `auto`, the engine keeps the backend chosen for its first
+  structured-output request.
+- `--sse-keep-alive-interval 15` sends an SSE comment line (`: keep-alive`) after 15 idle
+  seconds on streaming responses, so proxies with read timeouts keep the connection open
+  during queueing and prefill. The default, 0, sends none.
+- `TILELANG_CACHE_DIR` ([step 3](getting-started.md#3-environment)) keeps TileLang's compiled kernels on persistent storage.
+  Without persistent cache storage, a recreated container compiles them again.
+- Media limits and request budgets: see
+  [multimodal input limits](../.github/MULTIMODAL_INPUTS.md).
+
+## Sampling parameters
+
+The server takes default sampling values from the checkpoint's `generation_config.json`
+(`temperature` 1.0, `top_p` 0.95, `top_k` 20), but only for fields that a request
+leaves out. Many chat clients send their own `temperature` and `top_p`, so check the
+client settings. Qwen recommends:
+
+| Mode | `temperature` | `top_p` | `top_k` | `min_p` | `presence_penalty` | `repetition_penalty` |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Thinking (default) | 1.0 | 0.95 | 20 | 0.0 | 0.0 | 1.0 |
+| Non-thinking | 0.7 | 0.80 | 20 | 0.0 | 1.5 | 1.0 |
+
+- **Endless repetition:** Qwen suggests a `presence_penalty` between 0 and 2. Higher values
+  can mix languages and slightly lower quality.
+- `presence_penalty` can only be set per request. `--override-generation-config` sets
+  server defaults for `temperature`, `top_p`, `top_k`, `min_p`, `repetition_penalty` and
+  `max_new_tokens` only.
+- `top_k` and `min_p` are not standard OpenAI fields. With the OpenAI Python SDK, pass them
+  in `extra_body`.
+- These are the model vendor's values. The [speed tests](performance.md) use
+  `temperature=0`. The greedy `draft_sample_method` in the launch command applies only to
+  draft verification, not to these request settings.
+
+## YaRN and native context
+
+YaRN in vLLM is static: it applies the same scaling to every request, short or long.
+Qwen's model card notes that this can lower quality on shorter texts and advises
+enabling YaRN only when long contexts are needed. ReVolta's quality measurements
+compare against a BF16 reference that also uses YaRN, so they do not show this
+effect. If your requests stay within 262,144 tokens, do not make YaRN views: use
+`--max-model-len 262144` and serve the checkpoint directory itself (the materialized
+directory for Standard NVFP4 and FP4-Local). For the MTP draft, use the AWQ
+checkpoint's own `mtp-fp8/`, the scale pack's `mtp-fp8/` for Standard NVFP4, or
+FP4-Local's built-in MTP layer with no draft `model` field. The full acceptance ran
+with the 512K views.
+
+## System memory and storage
+
+**128 GiB RAM is recommended** for the default 48 GiB CPU KV pool. The AWQ PLE
+embedding table uses about 47.7 GiB of pinned host memory even with KV offload disabled.
+
+| Container memory limit | KV offload | Result |
+| --- | --- | --- |
+| 128 GiB | Off, or 48 GiB CPU pool | Full acceptance passed |
+| 80 GiB | Off | Memory stress passed; full acceptance not run |
+| 64 GiB | Off | Out of memory during startup |
+
+80 GiB is the smallest tested limit that passed memory stress, without full acceptance. With less than 128 GiB,
+reduce or disable the CPU KV pool.
+
+Image and video preprocessing also uses host memory, in the API server process. In one
+test, 8 concurrent requests with 8 different 2048-pixel images each used about 33 GB at
+peak. With a 32 GiB CPU KV pool on a 125 GiB host, the kernel then killed a worker. The
+same burst passed with a 16 GiB pool. If you expect bursts of large images or videos,
+use a smaller CPU KV pool or limit the image size and the number of concurrent requests.
+
+The AWQ checkpoint is 137 GB at the pinned revision:
+
+| Part | Format | Size | Where it runs |
+| --- | --- | ---: | --- |
+| Routed experts | 4-bit AWQ, group size 32 | 74.9 GB | GPU |
+| PLE n-gram tables | FP8 E4M3 | 51.3 GB | Pinned host memory |
+| Other weights (attention, embeddings, `lm_head`, vision tower, shared experts) | 16-bit | 10.9 GB | GPU |
+| Optional MTP draft (`mtp-fp8/`) | FP8 | 5.3 GB | GPU |
+
+Allow extra disk space for compilation caches and the optional disk KV tier.
+
+## See also
+
+- [Reference launch tutorial](getting-started.md)
+- [Checkpoint preparation](checkpoints.md)
+- [KV offload modes](kv-offload.md)