Add INT4 and FP4 lanes to the Ling-3.0-flash-VL cookbook (#38527)

docs/cookbook/autoregressive/InclusionAI/Ling-3.0-flash-VL.mdx

@@ -49,7 +49,7 @@ For how to launch the image, see [Install → Method 3: Using Docker](../../../d
 
 </Accordion>
 
-Pick your hardware to generate the launch command. The checkpoint ships in BF16 with an FP8 variant, and each GPU family carries one **Balanced** recipe sized so the ~250 GB of BF16 weights fit with KV-cache headroom: `--tp 4` on 288 GB-class (GB300, B300) and 141 GB-class (B200, H200) GPUs, and `--tp 8` on H100.
+Pick your hardware and quantization to generate the launch command. The checkpoint ships in BF16 with FP8, INT4, and FP4 variants, and each combination carries one **Balanced** recipe: BF16 uses `--tp 4` on 288 GB-class (GB300, B300) and 141 GB-class (B200, H200) GPUs and `--tp 8` on H100 (sized so the ~250 GB of BF16 weights fit with KV-cache headroom); FP8 uses `--tp 1` on GB300/B300/B200 and `--tp 2` on H200/H100; INT4 and FP4 use `--tp 1` everywhere. DGX Spark offers INT4 and FP4 only.
 
 import { Deployment } from "/src/snippets/_deployment.jsx";
 import { config } from "/src/snippets/configs/inclusionAI/ling-3.0-flash-vl.jsx";
@@ -74,6 +74,9 @@ It is a thinking model: the chat template turns chain-of-thought on by default a
 **Available Models:**
 
 - **BF16**: [inclusionAI/Ling-3.0-flash-VL](https://huggingface.co/inclusionAI/Ling-3.0-flash-VL) — ~125B total / ~5.1B active
+- **FP8**: [inclusionAI/Ling-3.0-flash-VL-FP8](https://huggingface.co/inclusionAI/Ling-3.0-flash-VL-FP8) — 128×128 block quantization
+- **INT4 (GPTQ)**: [inclusionAI/Ling-3.0-flash-VL-int4](https://huggingface.co/inclusionAI/Ling-3.0-flash-VL-int4) — compressed-tensors, group size 32
+- **FP4 (MXFP4)**: [inclusionAI/Ling-3.0-flash-VL-fp4](https://huggingface.co/inclusionAI/Ling-3.0-flash-VL-fp4) — e2m1 experts + FP8 e4m3 activations, Blackwell only
 
 **License:** MIT
 
@@ -91,7 +94,9 @@ It is a thinking model: the chat template turns chain-of-thought on by default a
 - Thinking is on by default (`enable_thinking` defaults to true in the chat template, and the `ling3` reasoning parser follows that default). Turn it off per request with `"chat_template_kwargs": {"enable_thinking": false}` (§3.3).
 - Native context is 128K, and the generated recipes default to the 256K YaRN variant (`--context-length 262144` plus the `rope_scaling` override and `SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1`), matching the model card. For the native 128K window, drop all three from the command.
 - Audio content parts are rejected before preprocessing; send images as `image_url` and clips as `video_url` content parts (§3).
-- The FP8 variant (`inclusionAI/Ling-3.0-flash-VL-FP8`) uses 128×128 block quantization, so plain TP must keep `768 ÷ tp` divisible by 128 (TP 1, 2, 3, or 6 — TP=4 fails to load). The FP8 recipes therefore pair TP with expert parallelism (`--tp 4 --ep 4`, `--tp 8 --ep 8`), which keeps experts whole per rank and is verified on 4×GB300; TP=2 without EP also works. The BF16 checkpoint can also be quantized online with `--quantization fp8`; measured accuracy and speed for both paths are in the benchmark cards below.
+- The FP8 variant (`inclusionAI/Ling-3.0-flash-VL-FP8`) uses 128×128 block quantization, so plain TP must keep `768 ÷ tp` divisible by 128 (TP 1, 2, 3, or 6 — TP=4 fails to load). FP8 recipes default to a single GPU where the ~126 GB weights fit (`--tp 1` on 288 GB-class and B200, verified on GB300) or `--tp 2` on tighter cards. As a manual alternative on four GPUs, `--tp 4 --ep 4` keeps experts whole per rank (measured on 4×GB300). The BF16 checkpoint can also be quantized online with `--quantization fp8`; measured accuracy and speed for both paths are in the benchmark cards below.
+- The INT4 variant (`inclusionAI/Ling-3.0-flash-VL-int4`) is a GPTQ 4-bit export (compressed-tensors, group size 32). It runs on all listed hardware with single-GPU (`--tp 1`) recipes; GB300, H200, and DGX Spark cells are verified, the rest carry the unverified badge.
+- The FP4 variant (`inclusionAI/Ling-3.0-flash-VL-fp4`) is an MXFP4 export (e2m1 routed-expert weights with e8m0 scales, group 32, plus FP8 e4m3 dynamic activations) and runs on Blackwell only (GB300, B300, B200, DGX Spark). The `flashinfer_mxfp4` MoE backend is selected automatically on the `lmsysorg/sglang:dev-Ling-3.0-flash-VL` image and on sglang with sgl-project/sglang#38526; on builds that already include the VL integration but predate the auto-selection, add `--moe-runner-backend flashinfer_mxfp4` explicitly. GB300 and DGX Spark cells are verified, the rest carry the unverified badge.
 
 ## 3. Advanced Usage