Add proper support for DGX spark with easy manager. Thank you NVIDIA!

manager/README.md

@@ -21,11 +21,11 @@ python3 -m manager doctor      # diagnose problems
 - **Hardware → spec mapping** is in [spec.py](spec.py). One universal torch
   pin (2.13.0 / torchvision 0.28.0 / torchaudio 2.11.0) on every platform:
   cu130 wheels when the driver supports CUDA 13 (cu126 fallback for older
-  drivers, refused outright on Blackwell GPUs which need cu130), same stack +
-  Python 3.11 + `dgx_requirements.txt` on DGX/Grace, PyPI wheels on Mac,
+  drivers, refused outright on Blackwell GPUs which need cu130), the same
+  stack on Linux aarch64 (DGX Spark / Grace), PyPI wheels on Mac,
   rocm7.1 (experimental) for AMD, `--cpu` to force a CPU install. **Torch
   pins there must be updated together with the README install instructions,
-  run_mac.zsh, and dgx_instructions.md.**
+  and run_mac.zsh.**
 - **Accelerators everywhere wheels exist**, via per-spec `extra_packages`
   (installed after requirements with `--upgrade` so they override pins) and
   `optional_packages` (installed one-by-one, warn-only on failure):

manager/detect.py

@@ -108,7 +108,4 @@ def detect():
 
     if IS_MAC:
         result["backend"] = "mps" if arch == "arm64" else "cpu"
-
-    # DGX OS / Grace (GB10, DGX Spark): NVIDIA GPU on aarch64 Linux
-    result["is_dgx"] = bool(result["os"] == "linux" and arch == "aarch64" and nvidia)
     return result

manager/env.py

@@ -81,6 +81,34 @@ def _venv_platform():
         return None
 
 
+def _venv_has_headers():
+    """True if the venv's base interpreter ships Python.h.
+
+    triton JIT-compiles a small C launcher (cuda_utils.c) on first GPU use and
+    needs the headers; a venv built on a distro Python without the -dev
+    package fails there with "Python.h: No such file or directory".
+    """
+    if not venv_exists():
+        return False
+    try:
+        out = subprocess.run(
+            [
+                venv_python(),
+                "-c",
+                "import os, sysconfig; print(os.path.isfile(os.path.join("
+                "sysconfig.get_paths()['include'], 'Python.h')))",
+            ],
+            stdout=subprocess.PIPE,
+            stderr=subprocess.DEVNULL,
+            timeout=30,
+            env=clean_env(),
+        )
+        # unknown -> assume fine rather than recreate on a flaky probe
+        return out.returncode != 0 or out.stdout.decode().strip() == "True"
+    except (OSError, subprocess.TimeoutExpired):
+        return True
+
+
 def _uv_python_platform(uv_python):
     """'win-arm64' / 'win-amd64' expected for a pinned uv interpreter request."""
     if not uv_python:
@@ -113,6 +141,21 @@ def ensure_venv(spec, dry_run=False):
             )
             import shutil
 
+            shutil.rmtree(venv_dir(), ignore_errors=True)
+        elif find_uv() and not _venv_has_headers():
+            if dry_run:
+                info(
+                    "[dry-run] venv Python has no C headers — would recreate "
+                    "it on a uv-managed Python."
+                )
+                return venv_python()
+            warn(
+                "Existing venv is built on a Python without C headers "
+                "(Python.h), which triton needs — recreating it on a "
+                "uv-managed Python (all packages will be reinstalled)."
+            )
+            import shutil
+
             shutil.rmtree(venv_dir(), ignore_errors=True)
         else:
             return venv_python()
@@ -134,7 +177,17 @@ def ensure_venv(spec, dry_run=False):
     if uv:
         info("Creating venv with uv (python %s) at %s" % (python_request, target))
         run(
-            [uv, "venv", target, "--python", python_request, "--seed"],
+            # --managed-python: never build on a distro Python, which may
+            # lack the C headers triton needs (see _venv_has_headers)
+            [
+                uv,
+                "venv",
+                target,
+                "--python",
+                python_request,
+                "--managed-python",
+                "--seed",
+            ],
             env=clean_env(),
         )
     else:
@@ -382,7 +435,6 @@ def requirements_hash(spec):
         for f in os.listdir(REPO_ROOT)
         if f.startswith("requirements") and f.endswith(".txt")
     ]
-    req_files.append(os.path.join(REPO_ROOT, "dgx_requirements.txt"))
     base = file_hash(req_files)
     import hashlib
 

manager/nodejs.py

@@ -26,7 +26,7 @@ from .util import (
 )
 
 NODE_DIR = os.path.join(REPO_ROOT, ".node")
-# Node 24 is the current LTS line and matches the dgx_instructions.md guidance.
+# Node 24 is the current LTS line.
 NODE_VERSION = "24.11.1"
 MIN_NODE_MAJOR = 20
 

manager/spec.py

@@ -306,16 +306,6 @@ def _cuda_spec(detection):
     arch = detection["arch"]
     flavor, notes = _cuda_flavor(detection)
     python_version = "3.12"
-    requirements = (
-        "dgx_requirements.txt" if detection.get("is_dgx") else "requirements.txt"
-    )
-    if detection.get("is_dgx"):
-        # the old "Python 3.11 on DGX OS" constraint was for conda/system
-        # installs; uv provisions 3.12 and all aarch64 cp312 wheels exist now
-        notes = notes + [
-            "DGX OS / Grace detected: using %s wheels and dgx_requirements.txt."
-            % flavor
-        ]
 
     extras = [TORCHCODEC]
     optional = [FLA]
@@ -340,7 +330,6 @@ def _cuda_spec(detection):
         TORCH,
         torch_index=PYTORCH_INDEX + flavor,
         python_version=python_version,
-        requirements_file=requirements,
         extra_packages=extras,
         optional_packages=optional,
         find_links=find_links,