[Kernel] Add native B12X linear, MoE, and causal attention backends

docs/features/quantization/b12x.md

@@ -0,0 +1,52 @@
+# B12X Backends
+
+[B12X](https://pypi.org/project/b12x/) provides optional CUDA kernels for
+NVIDIA SM120 and SM121 GPUs. Install the dependency with:
+
+```bash
+uv pip install "vllm[b12x]"
+```
+
+The B12X backends are opt-in. To select all supported B12X kernels explicitly:
+
+```bash
+vllm serve <model> \
+    --linear-backend b12x \
+    --moe-backend b12x \
+    --attention-backend B12X_ATTN
+```
+
+Only pass `--moe-backend b12x` for a compatible NVFP4 or MXFP4 MoE model. The
+linear and attention backends can be selected independently.
+
+B12X uses MXFP8 activations by default for MXFP4 MoE and the checkpoint's
+activation format for NVFP4 MoE. MXFP4 falls back to BF16 when its A8 path does
+not support the model configuration. Set `VLLM_B12X_MOE_FORCE_A16=1` to force
+BF16 activations for either weight format.
+
+For more targeted linear selection, leave `--linear-backend` at `auto` and
+enable B12X for one or both quantization families:
+
+```bash
+VLLM_USE_B12X_FP8_GEMM=1 VLLM_USE_B12X_FP4_GEMM=1 \
+    vllm serve <model>
+```
+
+An explicit non-B12X `--linear-backend` takes precedence over these environment
+variables.
+
+## Supported Configurations
+
+| Backend | Supported configurations |
+| ------- | ------------------------ |
+| Linear | Per-tensor FP8, 128x128 block FP8, MXFP8, NVFP4, and MXFP4 |
+| MoE | Tensor-parallel MXFP4 weights with BF16 or MXFP8 activations; NVFP4 weights with BF16, NVFP4, or MXFP8 activations |
+| Attention | Causal decoder MHA/MQA/GQA with BF16 model dtype; BF16, FP16, or FP8 E4M3 KV cache; head sizes 64, 128, 192, or 256; block size 64 or 128 |
+
+The attention backend supports decode, prefill, mixed batches, speculative
+verification, sliding-window attention, attention sinks, and CUDA graphs.
+
+The B12X MoE backend does not support expert parallelism, expert maps, EXL3, or
+NF3. The attention backend does not support MLA, non-causal attention, ALiBi,
+logits soft capping, or context parallelism. Dense W4A16 layers are not handled
+by B12X and continue to use another compatible backend such as Marlin.

setup.py

@@ -1291,6 +1291,7 @@ def add_vllm_package_data(filename: str) -> None:
         # only; also needs system GStreamer + libv4l (see docs).
         "deepstream": ["nvidia-deepstream-videodecode-cu13>=9.0.2"],
         "flashinfer": [],  # Kept for backwards compatibility
+        "b12x": ["b12x>=1.2.2"],
         # Optional deps for Helion kernel development
         # NOTE: When updating helion version, also update CI files:
         #   - .buildkite/test_areas/kernels.yaml