Consolidate DSpark, vision port, safetensors overlay, RCA fixes, and docs onto sm80

.dockerignore

@@ -1,35 +1,7 @@
-/.venv
-/build
-dist
-vllm/*.so
-vllm/vllm-rs
-
-# Byte-compiled / optimized / DLL files
-__pycache__/
-*.py[cod]
-*$py.class
-
-.mypy_cache
-
-# Distribution / packaging
-.Python
-/build/
-cmake-build-*/
-CMakeUserPresets.json
-develop-eggs/
-/dist/
-downloads/
-eggs/
-.eggs/
-lib/
-lib64/
-parts/
-sdist/
-var/
-wheels/
-share/python-wheels/
-*.egg-info/
-.installed.cfg
-*.egg
-MANIFEST
-rust/target/
+__pycache__
+*.pyc
+docker/
+notebooks/
+*.md
+receipts/
+**/receipts/

README.md

@@ -19,6 +19,52 @@ For events, please visit [vllm.ai/events](https://vllm.ai/events) to join us.
 
 ---
 
+## PixelML SM80 fork
+
+This fork carries NVIDIA Ampere (compute capability 8.0) fallback kernels
+for model families whose upstream vLLM support targets Hopper or newer.
+It exists for the PixelML `club-170hx` hardware group (4x NVIDIA CMP 170HX,
+GA100 silicon, 64 GiB HBM2e per card, no FP4/FP8 tensor cores, no NVLink).
+Full provenance, licenses, and per-branch status are in
+[`docs/SM80.md`](docs/SM80.md).
+
+### Supported models
+
+| Model | Branch | Status | Notes |
+| --- | --- | --- | --- |
+| DeepSeek-V4-Flash (text) | `sm80` | Serving, published | `ghcr.io/pixelml/club-170hx:vllm-deepseek-v4-sm80-20260902`; see `club-170hx` `docs/MODEL-STATUS.md` |
+| DeepSeek-V4-Flash-Vision-Exp | `sm80` | Serving, published; two known limits below | Reverse-ported vision path (Path 3), consolidated onto this branch; see `docs/SM80.md` and `PixelML/DeepSeek-V4-Flash-Vision-Exp-CMP-170HX` `docs/VISION-PORT.md` |
+| GLM-5.3-Flash | `glm53-sm80` | In progress (repo-only Phase A as of 2026-09-03) | See issue [#103](https://github.com/seanphan/pixelml/issues/103) for the port plan and inventory |
+| Qwen3.8-27B | stock vLLM (no fork changes needed) | Serving, published | No SM80 fallback required; see `club-170hx` `docs/MODEL-STATUS.md` |
+
+### Known limits (DeepSeek-V4-Flash-Vision-Exp)
+
+- **`c>=4` crash**: an EngineCore crash at concurrency 4 and above.
+  Root-caused and fixed on `sm80`; not yet re-verified on a GPU boot.
+  See [seanphan/pixelml#79](https://github.com/seanphan/pixelml/issues/79).
+- **131k-token crash**: prefill past 65,000 tokens crashes the engine.
+  Root-caused and fixed on `sm80`; not yet re-verified on a GPU boot.
+  See [seanphan/pixelml#79](https://github.com/seanphan/pixelml/issues/79).
+
+### Build instructions
+
+```bash
+git checkout sm80   # or glm53-sm80 once that branch has a bootable model
+TORCH_CUDA_ARCH_LIST=8.0 docker build -f docker/Dockerfile.sm80 -t <tag> .
+```
+
+A from-source build takes about 60 minutes on the `club-170hx` build host.
+See `club-170hx` `docs/DOCKER-IMAGE.md` for the exact published image
+digests and launch commands.
+
+### Links
+
+- Platform manual and benchmarks: [`PixelML/club-170hx`](https://github.com/PixelML/club-170hx)
+- DeepSeek-V4 vision port evidence: [`PixelML/DeepSeek-V4-Flash-Vision-Exp-CMP-170HX`](https://github.com/PixelML/DeepSeek-V4-Flash-Vision-Exp-CMP-170HX)
+- GLM-5.3-Flash tracking issue: [seanphan/pixelml#103](https://github.com/seanphan/pixelml/issues/103)
+
+---
+
 ## About
 
 vLLM is a fast and easy-to-use library for LLM inference and serving.

csrc/libtorch_stable/moe/moe_ops.h

@@ -28,7 +28,9 @@ void topk_softplus_sqrt(
     const std::optional<torch::stable::Tensor>& correction_bias,
     const std::optional<torch::stable::Tensor>& input_ids,
     const std::optional<torch::stable::Tensor>& tid2eid,
-    const std::optional<torch::stable::Tensor>& is_padding);
+    const std::optional<torch::stable::Tensor>& is_padding,
+    const std::optional<torch::stable::Tensor>& bias_vl,
+    int64_t image_sentinel_lo);
 
 void moe_sum(torch::stable::Tensor& input, torch::stable::Tensor& output,
              std::optional<torch::stable::Tensor> topk_ids,