[pull] main from NVIDIA:main

3rdparty/patches/msa_strided_paged_kv.patch

@@ -362,7 +362,7 @@ diff --git a/python/fmha_sm100/cute/test_sparse_atten.py b/python/fmha_sm100/cut
 index 21c777e..b5f078b 100644
 --- a/python/fmha_sm100/cute/test_sparse_atten.py
 +++ b/python/fmha_sm100/cute/test_sparse_atten.py
-@@ -61,6 +61,81 @@ DECODE_DIM = 128
+@@ -61,6 +61,78 @@ DECODE_DIM = 128
  DECODE_KV_TOKEN_SWEEP = tuple(2**exp for exp in range(3, 21))
 
 
@@ -387,14 +387,11 @@ index 21c777e..b5f078b 100644
 +
 +def test_prepare_paged_hnd_input_materializes_unpacked_tokens() -> None:
 +    pages, heads, page_size, head_dim = 5, 2, 128, 128
-+    storage = torch.empty(
-+        pages,
-+        heads,
-+        page_size * 2,
-+        head_dim,
-+        dtype=torch.float8_e4m3fn,
++    storage = torch.arange(
++        pages * heads * page_size * 2 * head_dim,
++        dtype=torch.int32,
 +        device="cuda",
-+    )
++    ).remainder(16).to(torch.float8_e4m3fn).reshape(pages, heads, page_size * 2, head_dim)
 +    view = storage[:, :, ::2, :]
 +
 +    prepared = sparse_interface._prepare_paged_hnd_input(view, page_size)
@@ -406,11 +403,11 @@ index 21c777e..b5f078b 100644
 +def test_prepare_paged_hnd_input_materializes_unaligned_outer_stride() -> None:
 +    pages, heads, page_size, head_dim = 5, 2, 128, 128
 +    outer_stride = heads * page_size * head_dim + 1
-+    storage = torch.empty(
++    storage = torch.arange(
 +        pages * outer_stride,
-+        dtype=torch.float8_e4m3fn,
++        dtype=torch.int32,
 +        device="cuda",
-+    )
++    ).remainder(16).to(torch.float8_e4m3fn)
 +    view = storage.as_strided(
 +        (pages, heads, page_size, head_dim),
 +        (outer_stride, page_size * head_dim, head_dim, 1),
@@ -444,7 +441,7 @@ index 21c777e..b5f078b 100644
  @contextmanager
  def _nvtx_range(message: str):
      torch.cuda.nvtx.range_push(message)
-@@ -1786,6 +1861,74 @@ def test_sparse_page_atten(
+@@ -1786,6 +1858,74 @@ def test_sparse_page_atten(
 
      _assert_forward_close(out, out_ref, out_pt.float(), lse, lse_ref)
 

docs/source/_ext/llmapi_config_telemetry.py

@@ -34,12 +34,16 @@
 for the wire schema.
 
 **No PII or free-form fields are captured.** LLM API configuration capture is
-*type-driven*: fields whose type is categorical (`Literal`/`Enum`/`bool`) or
-numeric (`int`/`float`), plus safe collections of those, are captured
-automatically. Free-form `str`/`Any`/`Path`/`dict`/`Callable` are never captured
-unless a field carries an explicit allowlist (`TelemetryField.categorical(...)`),
-and any field may opt out with `telemetry=False`. Every captured field is listed
-below; the runtime can capture nothing absent from this list.
+automatic for `bool`, `int`, finite `float`, `Literal`, `Enum`, supported unions,
+and homogeneous sequences. Unsafe scalar `str`, `Any`, and `object` branches
+require `TelemetryField.categorical(...)`; paths, mappings, callables, and
+unsupported structures always fail closed. Use `telemetry=False` to exclude a
+field. The runtime can capture nothing absent from the list below.
+
+`capture_policy` branches separated by `|` are tried independently; `enum[X]`
+requires the exact enum type `X`. The categorical domain lists tokens from
+`Literal`/`Enum` annotations or explicit `allowed_values`; it does not restrict
+`bool`, `int`, or `float` branches.
 
 If the manifest check fails, run `python3 scripts/generate_llm_args_golden_manifest.py`, then commit
 `tensorrt_llm/usage/llm_args_golden_manifest.json`; new fields require telemetry/privacy CODEOWNER approval.
@@ -55,20 +59,24 @@ def _escape(text: str) -> str:
     return text.replace("|", "\\|").replace("\n", " ")
 
 
-def _format_values(values: list[str]) -> str:
-    return ", ".join(f"`{_escape(v)}`" for v in values) if values else ""
+def _format_values(values: list[object]) -> str:
+    def format_value(value: object) -> str:
+        text = value if isinstance(value, str) else json.dumps(value)
+        return f"`{_escape(text)}`"
+
+    return ", ".join(format_value(value) for value in values)
 
 
 def _table(rows: list[dict]) -> str:
     lines = [
-        "| Captured key | Annotation | Kind | Converter | Allowed values |",
-        "|--------------|------------|------|-----------|----------------|",
+        "| Captured key | Capture policy | Kind | Categorical domain |",
+        "|--------------|----------------|------|--------------------|",
     ]
     for row in rows:
         lines.append(
-            f"| `{_escape(row['path'])}` | `{_escape(row['annotation'])}` | "
-            f"`{_escape(row['kind'])}` | {_escape(row['converter'])} | "
-            f"{_format_values(row['allowed_values'])} |"
+            f"| `{_escape(row['path'])}` | `{_escape(row['capture_policy'])}` | "
+            f"`{_escape(row['kind'])}` | "
+            f"{_format_values(row.get('allowed_values', []))} |"
         )
     return "\n".join(lines)
 

docs/source/_static/config_db.json

@@ -84,6 +84,18 @@
       "model_url": "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B-FP8",
       "scenario": "Max Throughput (Static EPLB, MTP3)"
     },
+    {
+      "command": "trtllm-serve Qwen/Qwen3.8-Flash-Next-FP8 --config ${TRTLLM_DIR}/examples/configs/curated/qwen3.8-flash-next.yaml",
+      "config_filename": "qwen3.8-flash-next.yaml",
+      "config_github_url": "https://github.com/NVIDIA/TensorRT-LLM/blob/main/examples/configs/curated/qwen3.8-flash-next.yaml",
+      "config_path": "examples/configs/curated/qwen3.8-flash-next.yaml",
+      "config_raw_url": "https://raw.githubusercontent.com/NVIDIA/TensorRT-LLM/main/examples/configs/curated/qwen3.8-flash-next.yaml",
+      "gpu_compatibility": "B200, B300, GB200, GB300",
+      "model": "Qwen/Qwen3.8-Flash-Next-FP8",
+      "model_display_name": "Qwen3.8-Flash-Next (FP8)",
+      "model_url": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8",
+      "scenario": "General deployment (MTP3)"
+    },
     {
       "command": "trtllm-serve nvidia/Qwen3.5-397B-A17B-NVFP4 --config ${TRTLLM_DIR}/examples/configs/curated/qwen3.5.yaml",
       "config_filename": "qwen3.5.yaml",
@@ -4278,6 +4290,10 @@
       "display_name": "Qwen3.8-2.4T-A95B (FP8)",
       "url": "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B-FP8"
     },
+    "Qwen/Qwen3.8-Flash-Next-FP8": {
+      "display_name": "Qwen3.8-Flash-Next (FP8)",
+      "url": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8"
+    },
     "deepseek-ai/DeepSeek-R1-0528": {
       "display_name": "DeepSeek-R1",
       "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528"