From fe5eee6a363591831253a4a079f56ff4b96f1913 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 1 Jul 2026 08:57:32 +0000
Subject: [PATCH] Lower llm-compressor floor to 0.6.0 so supported old torch
 still resolves

The >=0.8.0 floor conflicts with the torch this install pins in its constraints
file. Unsloth supports torch>=2.4, but llm-compressor 0.7.0+ require torch>=2.7
(0.10+ need >=2.9, 0.12+ need >=2.10). On a supported torch 2.4-2.6 box pip then
has no candidate in [0.8.0, 1.0) and FP8/FP4 export fails before quantization.

Lower the floor to 0.6.0 (its metadata only needs torch>=1.7), which never
conflicts with any supported torch. pip still prefers the newest compatible
release, so modern torch continues to get the latest 0.x (0.12.0). The <1.0
ceiling that blocks an inflated-version supply-chain jump is unchanged.

Add a regression test asserting the floor stays <= 0.6.0.
---
 .../saving/test_llm_compressor_install_pin.py | 20 +++++++++++++++++
 unsloth/save.py                               | 22 ++++++++++++-------
 2 files changed, 34 insertions(+), 8 deletions(-)

diff --git a/tests/saving/test_llm_compressor_install_pin.py b/tests/saving/test_llm_compressor_install_pin.py
index fc3f30fee0..89fdf52f84 100644
--- a/tests/saving/test_llm_compressor_install_pin.py
+++ b/tests/saving/test_llm_compressor_install_pin.py
@@ -54,6 +54,26 @@ def test_spec_is_a_bounded_pin() -> None:
     assert ">=" in spec and "<" in spec, f"spec must have lower and upper bounds, got {spec!r}"
 
 
+def test_floor_stays_compatible_with_supported_torch() -> None:
+    """The floor must not require a torch newer than the oldest torch Unsloth supports (>=2.4).
+
+    llm-compressor 0.7.0+ require torch>=2.7 (0.10+ >=2.9, 0.12+ >=2.10); only <=0.6.x allows
+    torch<2.7. Since install_llm_compressor() pins the current torch in the constraints file, a
+    floor above 0.6.0 leaves pip with no candidate on a supported torch 2.4-2.6 box and breaks
+    FP8/FP4 export. Keep the floor at or below 0.6.0.
+    """
+    from packaging.requirements import Requirement
+    from packaging.version import Version
+
+    req = Requirement(_spec_value())
+    lowers = [Version(s.version) for s in req.specifier if s.operator in (">=", "==", "~=")]
+    assert lowers, "spec must declare a lower bound"
+    assert max(lowers) <= Version("0.6.0"), (
+        f"floor {max(lowers)} requires a torch newer than Unsloth's minimum (2.4); "
+        "llm-compressor >0.6.0 needs torch>=2.7. Keep the floor <= 0.6.0."
+    )
+
+
 def test_install_command_uses_pinned_spec_not_bare_name() -> None:
     fn = _get_function("install_llm_compressor")
     # No argv list may pass the bare, unpinned package literal "llmcompressor".
diff --git a/unsloth/save.py b/unsloth/save.py
index 4355d77f6d..b9a7cb1dfc 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -1367,14 +1367,20 @@ def install_python_non_blocking(packages = []):
 # bound, `pip install llmcompressor` resolves to whatever the configured index offers, so a
 # compromised, dependency-confused, or inflated-version ("999.0.0") release could be pulled and
 # executed under the Unsloth process at install/import time; the "<1.0" ceiling blocks that jump.
-# The floor keeps the oneshot / QuantizationModifier API this uses. Crucially the range still lets
-# pip pick up new 0.x releases, which is where support for brand-new architectures lands (the gate
-# below can require a newer llm-compressor for newer schemes/models) -- an exact pin would break
-# exporting new models. Bump the ceiling deliberately when llm-compressor reaches 1.0 so a new
-# major is vetted before it is auto-installed. An already-installed newer llm-compressor is used
-# as-is (the import below short-circuits), so this only constrains the auto-install, never what
-# the user installed themselves.
-_LLM_COMPRESSOR_SPEC = "llmcompressor>=0.8.0,<1.0"
+# The range still lets pip pick up new 0.x releases, which is where support for brand-new
+# architectures lands (the gate below can require a newer llm-compressor for newer schemes/models)
+# -- an exact pin would break exporting new models. Bump the ceiling deliberately when
+# llm-compressor reaches 1.0 so a new major is vetted before it is auto-installed.
+#
+# The floor is 0.6.0 (its metadata only needs torch>=1.7) so it never conflicts with the torch this
+# install pins in the constraints file below: Unsloth supports torch>=2.4, but llm-compressor
+# 0.7.0+ require torch>=2.7 (0.10+ need >=2.9, 0.12+ need >=2.10). A higher floor would leave pip
+# with no candidate on a supported torch 2.4-2.6 box and break FP8/FP4 export before quantization.
+# pip still prefers the newest compatible release, so modern torch gets the latest 0.x anyway.
+#
+# An already-installed newer llm-compressor is used as-is (the import below short-circuits), so this
+# only constrains the auto-install, never what the user installed themselves.
+_LLM_COMPRESSOR_SPEC = "llmcompressor>=0.6.0,<1.0"
 
 
 def install_llm_compressor():