Add flash-linear-attention package for hardware that supports it.

manager/env.py

@@ -441,15 +441,23 @@ def _stale_git_pins(spec):
     return [name for name, line in current.items() if stored.get(name) != line]
 
 
-def _optional_import_name(pkg):
-    """Importable module name for an optional package spec or wheel URL."""
+# optional packages whose import name differs from the distribution name
+_IMPORT_ALIASES = {"flash_linear_attention": "fla"}
+# companion dists to remove on rollback (fla-core provides the `fla` module
+# itself; leaving it behind would keep the broken import resolvable)
+_ROLLBACK_EXTRAS = {"flash_linear_attention": ["fla-core"]}
+
+
+def _optional_names(pkg):
+    """(distribution, import) names for an optional package spec or wheel URL."""
     if "://" in pkg:
         name = os.path.basename(pkg).split("-")[0]
     else:
         name = pkg
         for sep in ("==", ">=", "<=", "<", ">", "["):
             name = name.split(sep)[0]
-    return name.strip().replace("-", "_")
+    name = name.strip().replace("-", "_")
+    return name, _IMPORT_ALIASES.get(name, name)
 
 
 def _venv_import_ok(module_name):
@@ -509,13 +517,15 @@ def ensure_requirements(spec, dry_run=False, force=False):
     # actually intend to ship, so repair it first if something got through
     _verify_torch(spec, dry_run=dry_run)
     # accelerators (flash-attn, NATTEN, ...): install one-by-one, warn on
-    # failure — training works without them, so never fail the whole install
+    # failure — training works without them, so never fail the whole install.
+    # No --upgrade here: every optional spec is an exact pin or wheel URL, and
+    # uv's -U eagerly upgrades the whole dependency closure, blowing past
+    # requirements.txt pins (numpy/transformers) that only the requirements
+    # pass enforces.
     for pkg in _filter_extras(spec.optional_packages):
         label = os.path.basename(pkg) if "://" in pkg else pkg
         info("Installing optional accelerator: %s" % label)
-        code = _pip_install(
-            [pkg] + find_links, dry_run=dry_run, upgrade=True, check=False
-        )
+        code = _pip_install([pkg] + find_links, dry_run=dry_run, check=False)
         if code != 0:
             warn("Optional package failed to install (continuing): %s" % label)
             continue
@@ -524,13 +534,14 @@ def ensure_requirements(spec, dry_run=False, force=False):
         # prebuilt accelerator wheels are sometimes built against a torch
         # nightly and fail to load against the release ABI — verify the
         # import and roll back rather than leaving a broken wheel installed
-        name = _optional_import_name(pkg)
-        if not _venv_import_ok(name):
+        dist_name, import_name = _optional_names(pkg)
+        if not _venv_import_ok(import_name):
             warn(
                 "%s installed but fails to import against this torch build — "
-                "removing it (training falls back to native attention)." % name
+                "removing it (training falls back to native attention)."
+                % import_name
             )
-            _pip_uninstall([name])
+            _pip_uninstall([dist_name] + _ROLLBACK_EXTRAS.get(dist_name, []))
     # packages whose dependency metadata is unsatisfiable on this platform but
     # which work fine without it (e.g. tensorboard's grpcio on win_arm64)
     for pkg in spec.no_deps_packages:

manager/spec.py

@@ -21,6 +21,11 @@ kernels smoke-tested on an RTX 5090 / sm120 with torch 2.13.0+cu130):
   {cp310..cp314} x {linux x86_64, linux aarch64}. No Windows/mac wheels.
 - triton: bundled with torch on Linux (incl. aarch64; torch 2.13 bundles
   triton 3.7.1); triton-windows 3.7.x matches on Windows; nothing for MPS.
+- flash-linear-attention 0.5.2: pure-Python (py3-none-any) Triton kernels —
+  installs anywhere, needs triton>=3.3 at runtime. Installed bare (no backend
+  extra) so it never pulls its own torch/triton over our pinned stack; usable
+  on every platform with a working triton (CUDA linux/windows, Spark, ROCm),
+  not on MPS/CPU.
 - Windows-on-ARM (verified 2026-07): there are NO win_arm64 wheels for any of
   the CUDA stack — torch cu130, triton-windows, flash-attn and Prisma's node
   engine are all x64-only. The supported configuration is therefore the x64
@@ -59,6 +64,10 @@ TRITON_WINDOWS = "triton-windows>=3.7,<3.8"
 NATTEN_VERSION = "0.21.7"
 NATTEN_FIND_LINKS = "https://whl.natten.org"
 
+# pure-Python triton kernels; bare install (no [cuda]/[rocm] extra) on purpose —
+# the extras only add torch/triton pins we already manage per-platform
+FLA = "flash-linear-attention==0.5.2"
+
 FLASH_ATTN_VERSION = "2.8.3"
 _FA_BASE = (
     "https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/"
@@ -309,7 +318,7 @@ def _cuda_spec(detection):
         ]
 
     extras = [TORCHCODEC]
-    optional = []
+    optional = [FLA]
     find_links = []
 
     # Windows-on-ARM runs the x64 wheel stack (see module docstring), so wheel
@@ -370,7 +379,11 @@ def _make_spark_spec(detection, wheels_source, dll_dirs):
             "triton==3.8.0+git8743423b",
         ],
         # import-verified with rollback, like accelerators on other platforms
-        optional_packages=["flash-attn==2.8.3+cu134torch2.14", "natten==0.21.7"],
+        optional_packages=[
+            "flash-attn==2.8.3+cu134torch2.14",
+            "natten==0.21.7",
+            FLA,
+        ],
         find_links=[wheels_source],
         notes=[
             "RTX Spark native mode: win_arm64 CUDA %s stack from the "
@@ -423,7 +436,10 @@ def _build_spec(detection, allow_cpu=False):
     os_name = detection["os"]
 
     if os_name == "mac":
-        notes = ["flash-attn / NATTEN / triton are unavailable on macOS."]
+        notes = [
+            "flash-attn / NATTEN / triton / flash-linear-attention are "
+            "unavailable on macOS."
+        ]
         if detection["backend"] != "mps":
             notes.append("Intel Mac detected — training will be extremely slow.")
         return EnvSpec(
@@ -443,6 +459,8 @@ def _build_spec(detection, allow_cpu=False):
             TORCH,
             torch_index=PYTORCH_INDEX + "rocm7.1",
             extra_packages=[TORCHCODEC],
+            # runs on ROCm via the triton bundled with rocm torch
+            optional_packages=[FLA],
             notes=[
                 "AMD ROCm support is experimental and largely untested.",
                 "flash-attn / NATTEN prebuilt wheels are unavailable for ROCm.",