[None][fix] Proper conditional compilation of sm10x cubins (#10839)

Signed-off-by: Yuan Tong <13075180+tongyuantongyu@users.noreply.github.com>
2026-02-04 18:21:52 +08:00 · 2026-01-28 10:17:51 +08:00 · 2026-01-28 10:17:51 +08:00 · 30348b2753
commit 30348b2753
parent c26a8f764c
2902 changed files with 27189 additions and 24336 deletions
--- a/.cmake-format.json
+++ b/.cmake-format.json
@ -0,0 +1,284 @@
+{
+  "_help_parse": "Options affecting listfile parsing",
+  "parse": {
+    "_help_additional_commands": [
+      "Specify structure for custom cmake functions"
+    ],
+    "additional_commands": {
+      "filter_source_cuda_architectures": {
+        "flags": [
+          "IMPLICIT_FAMILY"
+        ],
+        "kwargs": {
+          "SOURCE_LIST": "1",
+          "TARGET": "1",
+          "ARCHS": "+"
+        }
+      }
+    },
+    "_help_vartags": [
+      "Specify variable tags."
+    ],
+    "vartags": [],
+    "_help_proptags": [
+      "Specify property tags."
+    ],
+    "proptags": []
+  },
+  "_help_format": "Options affecting formatting.",
+  "format": {
+    "_help_line_width": [
+      "How wide to allow formatted cmake files"
+    ],
+    "line_width": 80,
+    "_help_tab_size": [
+      "How many spaces to tab for indent"
+    ],
+    "tab_size": 2,
+    "_help_max_subgroups_hwrap": [
+      "If an argument group contains more than this many sub-groups",
+      "(parg or kwarg groups) then force it to a vertical layout."
+    ],
+    "max_subgroups_hwrap": 2,
+    "_help_max_pargs_hwrap": [
+      "If a positional argument group contains more than this many",
+      "arguments, then force it to a vertical layout."
+    ],
+    "max_pargs_hwrap": 6,
+    "_help_max_rows_cmdline": [
+      "If a cmdline positional group consumes more than this many",
+      "lines without nesting, then invalidate the layout (and nest)"
+    ],
+    "max_rows_cmdline": 2,
+    "_help_separate_ctrl_name_with_space": [
+      "If true, separate flow control names from their parentheses",
+      "with a space"
+    ],
+    "separate_ctrl_name_with_space": false,
+    "_help_separate_fn_name_with_space": [
+      "If true, separate function names from parentheses with a",
+      "space"
+    ],
+    "separate_fn_name_with_space": false,
+    "_help_dangle_parens": [
+      "If a statement is wrapped to more than one line, than dangle",
+      "the closing parenthesis on its own line."
+    ],
+    "dangle_parens": false,
+    "_help_dangle_align": [
+      "If the trailing parenthesis must be 'dangled' on its on",
+      "line, then align it to this reference: `prefix`: the start",
+      "of the statement,  `prefix-indent`: the start of the",
+      "statement, plus one indentation  level, `child`: align to",
+      "the column of the arguments"
+    ],
+    "dangle_align": "prefix",
+    "_help_min_prefix_chars": [
+      "If the statement spelling length (including space and",
+      "parenthesis) is smaller than this amount, then force reject",
+      "nested layouts."
+    ],
+    "min_prefix_chars": 4,
+    "_help_max_prefix_chars": [
+      "If the statement spelling length (including space and",
+      "parenthesis) is larger than the tab width by more than this",
+      "amount, then force reject un-nested layouts."
+    ],
+    "max_prefix_chars": 10,
+    "_help_max_lines_hwrap": [
+      "If a candidate layout is wrapped horizontally but it exceeds",
+      "this many lines, then reject the layout."
+    ],
+    "max_lines_hwrap": 2,
+    "_help_line_ending": [
+      "What style line endings to use in the output."
+    ],
+    "line_ending": "unix",
+    "_help_command_case": [
+      "Format command names consistently as 'lower' or 'upper' case"
+    ],
+    "command_case": "canonical",
+    "_help_keyword_case": [
+      "Format keywords consistently as 'lower' or 'upper' case"
+    ],
+    "keyword_case": "unchanged",
+    "_help_always_wrap": [
+      "A list of command names which should always be wrapped"
+    ],
+    "always_wrap": [],
+    "_help_enable_sort": [
+      "If true, the argument lists which are known to be sortable",
+      "will be sorted lexicographicall"
+    ],
+    "enable_sort": true,
+    "_help_autosort": [
+      "If true, the parsers may infer whether or not an argument",
+      "list is sortable (without annotation)."
+    ],
+    "autosort": false,
+    "_help_require_valid_layout": [
+      "By default, if cmake-format cannot successfully fit",
+      "everything into the desired linewidth it will apply the",
+      "last, most agressive attempt that it made. If this flag is",
+      "True, however, cmake-format will print error, exit with non-",
+      "zero status code, and write-out nothing"
+    ],
+    "require_valid_layout": false,
+    "_help_layout_passes": [
+      "A dictionary mapping layout nodes to a list of wrap",
+      "decisions. See the documentation for more information."
+    ],
+    "layout_passes": {}
+  },
+  "_help_markup": "Options affecting comment reflow and formatting.",
+  "markup": {
+    "_help_bullet_char": [
+      "What character to use for bulleted lists"
+    ],
+    "bullet_char": "*",
+    "_help_enum_char": [
+      "What character to use as punctuation after numerals in an",
+      "enumerated list"
+    ],
+    "enum_char": ".",
+    "_help_first_comment_is_literal": [
+      "If comment markup is enabled, don't reflow the first comment",
+      "block in each listfile. Use this to preserve formatting of",
+      "your copyright/license statements."
+    ],
+    "first_comment_is_literal": false,
+    "_help_literal_comment_pattern": [
+      "If comment markup is enabled, don't reflow any comment block",
+      "which matches this (regex) pattern. Default is `None`",
+      "(disabled)."
+    ],
+    "literal_comment_pattern": null,
+    "_help_fence_pattern": [
+      "Regular expression to match preformat fences in comments",
+      "default= ``r'^\\s*([`~]{3}[`~]*)(.*)$'``"
+    ],
+    "fence_pattern": "^\\s*([`~]{3}[`~]*)(.*)$",
+    "_help_ruler_pattern": [
+      "Regular expression to match rulers in comments default=",
+      "``r'^\\s*[^\\w\\s]{3}.*[^\\w\\s]{3}$'``"
+    ],
+    "ruler_pattern": "^\\s*[^\\w\\s]{3}.*[^\\w\\s]{3}$",
+    "_help_explicit_trailing_pattern": [
+      "If a comment line matches starts with this pattern then it",
+      "is explicitly a trailing comment for the preceeding",
+      "argument. Default is '#<'"
+    ],
+    "explicit_trailing_pattern": "#<",
+    "_help_hashruler_min_length": [
+      "If a comment line starts with at least this many consecutive",
+      "hash characters, then don't lstrip() them off. This allows",
+      "for lazy hash rulers where the first hash char is not",
+      "separated by space"
+    ],
+    "hashruler_min_length": 10,
+    "_help_canonicalize_hashrulers": [
+      "If true, then insert a space between the first hash char and",
+      "remaining hash chars in a hash ruler, and normalize its",
+      "length to fill the column"
+    ],
+    "canonicalize_hashrulers": true,
+    "_help_enable_markup": [
+      "enable comment markup parsing and reflow"
+    ],
+    "enable_markup": true
+  },
+  "_help_lint": "Options affecting the linter",
+  "lint": {
+    "_help_disabled_codes": [
+      "a list of lint codes to disable"
+    ],
+    "disabled_codes": [],
+    "_help_function_pattern": [
+      "regular expression pattern describing valid function names"
+    ],
+    "function_pattern": "[0-9a-z_]+",
+    "_help_macro_pattern": [
+      "regular expression pattern describing valid macro names"
+    ],
+    "macro_pattern": "[0-9A-Z_]+",
+    "_help_global_var_pattern": [
+      "regular expression pattern describing valid names for",
+      "variables with global (cache) scope"
+    ],
+    "global_var_pattern": "[A-Z][0-9A-Z_]+",
+    "_help_internal_var_pattern": [
+      "regular expression pattern describing valid names for",
+      "variables with global scope (but internal semantic)"
+    ],
+    "internal_var_pattern": "_[A-Z][0-9A-Z_]+",
+    "_help_local_var_pattern": [
+      "regular expression pattern describing valid names for",
+      "variables with local scope"
+    ],
+    "local_var_pattern": "[a-z][a-z0-9_]+",
+    "_help_private_var_pattern": [
+      "regular expression pattern describing valid names for",
+      "privatedirectory variables"
+    ],
+    "private_var_pattern": "_[0-9a-z_]+",
+    "_help_public_var_pattern": [
+      "regular expression pattern describing valid names for public",
+      "directory variables"
+    ],
+    "public_var_pattern": "[A-Z][0-9A-Z_]+",
+    "_help_argument_var_pattern": [
+      "regular expression pattern describing valid names for",
+      "function/macro arguments and loop variables."
+    ],
+    "argument_var_pattern": "[a-z][a-z0-9_]+",
+    "_help_keyword_pattern": [
+      "regular expression pattern describing valid names for",
+      "keywords used in functions or macros"
+    ],
+    "keyword_pattern": "[A-Z][0-9A-Z_]+",
+    "_help_max_conditionals_custom_parser": [
+      "In the heuristic for C0201, how many conditionals to match",
+      "within a loop in before considering the loop a parser."
+    ],
+    "max_conditionals_custom_parser": 2,
+    "_help_min_statement_spacing": [
+      "Require at least this many newlines between statements"
+    ],
+    "min_statement_spacing": 1,
+    "_help_max_statement_spacing": [
+      "Require no more than this many newlines between statements"
+    ],
+    "max_statement_spacing": 2,
+    "max_returns": 6,
+    "max_branches": 12,
+    "max_arguments": 5,
+    "max_localvars": 15,
+    "max_statements": 50
+  },
+  "_help_encode": "Options affecting file encoding",
+  "encode": {
+    "_help_emit_byteorder_mark": [
+      "If true, emit the unicode byte-order mark (BOM) at the start",
+      "of the file"
+    ],
+    "emit_byteorder_mark": false,
+    "_help_input_encoding": [
+      "Specify the encoding of the input file. Defaults to utf-8"
+    ],
+    "input_encoding": "utf-8",
+    "_help_output_encoding": [
+      "Specify the encoding of the output file. Defaults to utf-8.",
+      "Note that cmake only claims to support utf-8 so be careful",
+      "when using anything else"
+    ],
+    "output_encoding": "utf-8"
+  },
+  "_help_misc": "Miscellaneous configurations options.",
+  "misc": {
+    "_help_per_command": [
+      "A dictionary containing any per-command configuration",
+      "overrides. Currently only `command_case` is supported."
+    ],
+    "per_command": {}
+  }
+}
--- a/cpp/cmake/modules/cuda_configuration.cmake
+++ b/cpp/cmake/modules/cuda_configuration.cmake
@ -15,10 +15,207 @@
 # the License.
 #

+#[=======================================================================[.rst:
+CudaConfiguration
+-----------------
+
+CUDA compiler and architecture configuration for TensorRT-LLM.
+
+This module provides functions and macros to configure the CUDA compiler,
+manage CUDA architectures, and filter source files based on target
+architectures. It is tailored to meet TensorRT-LLM's specific requirements
+for optimized kernel compilation across multiple GPU generations.
+
+Macros
+^^^^^^
+
+.. command:: setup_cuda_compiler
+
+  Detects and validates the CUDA compiler::
+
+    setup_cuda_compiler()
+
+  This macro determines the CUDA compiler version before enabling the CUDA
+  language extension. It requires CUDA version 11.2 or later.
+
+  The macro sets ``CMAKE_CUDA_COMPILER_VERSION`` upon successful detection.
+
+Functions
+^^^^^^^^^
+
+.. command:: setup_cuda_architectures
+
+  Initializes and normalizes ``CMAKE_CUDA_ARCHITECTURES``::
+
+    setup_cuda_architectures()
+
+  This function processes the ``CMAKE_CUDA_ARCHITECTURES`` variable and
+  configures architecture-specific compilation settings. This function should
+  be called after enabling the CUDA language extension.
+
+  **Special Values for CMAKE_CUDA_ARCHITECTURES:**
+
+  ``native``
+    Resolves to the highest available architecture on the system.
+    Falls back to ``all`` if detection fails.
+
+  ``all`` or unset
+    Resolves to architectures TensorRT-LLM is optimized for and the
+    compiler supports (80, 86, 89, 90, 100, 103, 120 depending on CUDA version).
+
+  ``all-major``
+    Unsupported. Results in a fatal error.
+
+  **Architecture Processing:**
+
+  * PTX is never included in the result binary (``-virtual`` rejected).
+  * The ``-real`` suffix is automatically added to exclude PTX.
+  * Accelerated targets (``-a`` suffix) are used for SM 90+.
+  * On CUDA 12.9+, family targets (``-f`` suffix) are used for SM 100+.
+
+  **Output Variables (set in parent scope):**
+
+  ``CMAKE_CUDA_ARCHITECTURES``
+    Normalized list with appropriate suffixes (e.g., ``80-real``, ``90a-real``,
+    ``100f-real``).
+
+  ``CMAKE_CUDA_ARCHITECTURES_ORIG``
+    Original list of enabled architectures without suffixes.
+
+  ``CMAKE_CUDA_ARCHITECTURES_FAMILIES``
+    List of family architectures (e.g., ``100f``, ``120f``).
+
+  ``CMAKE_CUDA_ARCHITECTURES_HAS_FAMILIES``
+    Boolean indicating if family targets are supported.
+
+  ``CMAKE_CUDA_MIN_ARCHITECTURE_HAS_ACCEL``
+    Minimum architecture supporting accelerated (``-a``) suffix.
+
+  ``CMAKE_CUDA_MIN_ARCHITECTURE_HAS_FAMILY``
+    Minimum architecture supporting family (``-f``) suffix.
+
+.. command:: add_cuda_architectures
+
+  Appends CUDA architectures to an existing target::
+
+    add_cuda_architectures(<target> <arch1> [<arch2> ...])
+
+  Adds the specified architectures to ``<target>``'s ``CUDA_ARCHITECTURES``
+  property. The ``-a`` suffix is automatically added for supported
+  architectures. Architectures are only added if they were explicitly
+  requested by the user in ``CMAKE_CUDA_ARCHITECTURES_ORIG``.
+
+.. command:: set_cuda_architectures
+
+  Sets CUDA architectures for a target::
+
+    set_cuda_architectures(<target> <arch1> [<arch2> ...])
+
+  Replaces the ``CUDA_ARCHITECTURES`` property of ``<target>`` with the
+  specified architectures.
+
+  **Architecture Specification:**
+
+  * Architectures may include the ``f`` suffix for family-conditional
+    compilation (e.g., ``100f``).
+  * Non-family architectures are only added if explicitly requested.
+  * Family architectures are only added if requested architectures would
+    enable compilation for that family.
+
+  If no architectures are enabled for the target, it compiles with
+  ``PLACEHOLDER_KERNELS`` macro defined. The kernel source shall compile
+  with any architecture if ``PLACEHOLDER_KERNELS`` macro is defined.
+
+.. command:: filter_source_cuda_architectures
+
+  Filters source files based on enabled CUDA architectures::
+
+    filter_source_cuda_architectures(
+      SOURCE_LIST <variable>
+      TARGET <target>
+      ARCHS <arch1> [<arch2> ...]
+      [IMPLICIT_FAMILY]
+    )
+
+  Removes source files targeting disabled CUDA architectures from the
+  source list. Files are matched by patterns like ``sm80``, ``sm_80``,
+  ``SM80``, etc. in their filenames (for ``.cu`` and ``cubin.cpp`` files).
+
+  ``SOURCE_LIST <variable>``
+    Name of the variable containing the list of source files.
+    Modified in place to remove filtered files.
+
+  ``TARGET <target>``
+    Target to add compile definitions to. If the target does not exist,
+    an INTERFACE library will be created.
+
+  ``ARCHS <arch1> [<arch2> ...]``
+    List of architectures to check. May include ``f`` suffix.
+
+  ``IMPLICIT_FAMILY``
+    When set, treats architectures >= ``CMAKE_CUDA_MIN_ARCHITECTURE_HAS_FAMILY``
+    as implicitly family-enabled.
+
+  **Defined Macros:**
+
+  For each filtered architecture, a compile definition ``EXCLUDE_SM_<ARCH>``
+  (or ``EXCLUDE_SM_<ARCH>F`` for family architectures) is added to ``<target>``.
+
+Example
+^^^^^^^
+
+.. code-block:: cmake
+
+  include(cuda_configuration)
+
+  # Setup compiler and detect version
+  setup_cuda_compiler()
+
+  # enable_language, or project(project_name LANGUAGES CUDA)
+  # must be called after setup_cuda_compiler() and before
+  # setup_cuda_architectures()
+  enable_language(CUDA)
+
+  # Configure architectures (uses CMAKE_CUDA_ARCHITECTURES if set)
+  setup_cuda_architectures()
+
+  # Add additional architecture to compile for, if it is beneficial.
+  # e.g. Utilizing native FP8 support available in sm89 (Ada)
+  # but not in sm86 (Ampere)
+  # Note: The kernel source must still compiles for all the architectures,
+  # by using less performant implementation.
+  add_library(my_kernels_fp8 STATIC kernels.cu)
+  add_cuda_architectures(my_kernels_fp8 89)
+
+  # Set specific architecture this source should compile for.
+  # e.g. Kernels using WGMMA instructions
+  # Note: The kernel source must still compiles for other architectures when
+  # ``PLACEHOLDER_KERNELS`` macro is defined.
+  add_library(my_kernels_sm90_only STATIC kernels.cu)
+  set_cuda_architectures(my_kernels_sm90_only 90)
+
+  # Filter sources for disabled architectures
+  set(KERNEL_SOURCES
+    kernel_sm80.cubin.cpp
+    kernel_sm90.cubin.cpp
+    kernel_sm100.cubin.cpp
+  )
+  filter_source_cuda_architectures(
+    SOURCE_LIST KERNEL_SOURCES
+    TARGET my_kernel_interface
+    ARCHS 80 90 100
+  )
+  # ``my_kernel_interface`` target is created with definitions to exclude
+  # disabled architectures.
+
+#]=======================================================================]
+
+#[[
+Determine CUDA version before enabling the language extension
+check_language(CUDA) clears CMAKE_CUDA_HOST_COMPILER if CMAKE_CUDA_COMPILER
+is not set
+#]]
 macro(setup_cuda_compiler)
-  # Determine CUDA version before enabling the language extension
-  # check_language(CUDA) clears CMAKE_CUDA_HOST_COMPILER if CMAKE_CUDA_COMPILER
-  # is not set
  include(CheckLanguage)
  if(NOT CMAKE_CUDA_COMPILER AND CMAKE_CUDA_HOST_COMPILER)
    set(CMAKE_CUDA_HOST_COMPILER_BACKUP ${CMAKE_CUDA_HOST_COMPILER})
@ -70,25 +267,28 @@ macro(setup_cuda_compiler)
  endif()
 endmacro()

-function(setup_cuda_architectures)
-  # cmake-format: off
-  # Initialize and normalize CMAKE_CUDA_ARCHITECTURES.
-  # Special values:
-  # * `native` is resolved to HIGHEST available architecture.
-  #   * Fallback to `all` if detection failed.
-  # * `all`/unset is resolved to a set of architectures we optimized for and compiler supports.
-  # * `all-major` is unsupported.
-  # Numerical architectures:
-  # * PTX is never included in result binary.
-  #   * `*-virtual` architectures are therefore rejected.
-  #   * `-real` suffix is automatically added to exclude PTX.
-  # * Always use accelerated (`-a` suffix) target for supported architectures.
-  # * On CUDA 12.9 or newer, family (`-f` suffix) target will be used for supported architectures to reduce number of
-  #   targets to compile for.
-  #   * Extra architectures can be requested via add_cuda_architectures
-  #     for kernels that benefit from arch specific features.
-  # cmake-format: on
+#[[
+Initialize and normalize CMAKE_CUDA_ARCHITECTURES.

+Special values:
+
+* `native` is resolved to HIGHEST available architecture.
+  * Fallback to `all` if detection failed.
+* `all`/unset is resolved to a set of architectures we optimized for and compiler supports.
+* `all-major` is unsupported.
+
+Numerical architectures:
+
+* PTX is never included in result binary.
+  * `*-virtual` architectures are therefore rejected.
+  * `-real` suffix is automatically added to exclude PTX.
+* Always use accelerated (`-a` suffix) target for supported architectures.
+* On CUDA 12.9 or newer, family (`-f` suffix) target will be used for supported architectures to reduce number of
+  targets to compile for.
+  * Extra architectures can be requested via add_cuda_architectures
+    for kernels that benefit from arch specific features.
+#]]
+function(setup_cuda_architectures)
  set(CMAKE_CUDA_ARCHITECTURES_RAW ${CMAKE_CUDA_ARCHITECTURES})
  if(CMAKE_CUDA_ARCHITECTURES_RAW STREQUAL "native")
    # Detect highest available compute capability
@ -138,9 +338,6 @@ function(setup_cuda_architectures)
        message(FATAL_ERROR "Unrecognized CUDA architecture: ${CUDA_ARCH}")
      endif()
    endforeach()
-    if("103" IN_LIST CMAKE_CUDA_ARCHITECTURES_CLEAN)
-      list(APPEND CMAKE_CUDA_ARCHITECTURES_CLEAN "100")
-    endif()
    list(REMOVE_DUPLICATES CMAKE_CUDA_ARCHITECTURES_CLEAN)
    set(CMAKE_CUDA_ARCHITECTURES_RAW ${CMAKE_CUDA_ARCHITECTURES_CLEAN})
  endif()
@ -182,22 +379,29 @@ function(setup_cuda_architectures)
  endforeach()

  # -a suffix supported from Hopper (90)
-  set(MIN_ARCHITECTURE_HAS_ACCEL 90)
+  set(CMAKE_CUDA_MIN_ARCHITECTURE_HAS_ACCEL 90)
+  set(CMAKE_CUDA_MIN_ARCHITECTURE_HAS_ACCEL
+      ${CMAKE_CUDA_MIN_ARCHITECTURE_HAS_ACCEL}
+      PARENT_SCOPE)
  # -f suffix supported from Blackwell (100) starting from CUDA 12.9.
  if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL "12.9")
-    set(MIN_ARCHITECTURE_HAS_FAMILY 100)
+    set(CMAKE_CUDA_MIN_ARCHITECTURE_HAS_FAMILY 100)
    set(CMAKE_CUDA_ARCHITECTURES_HAS_FAMILIES
        ON
        PARENT_SCOPE)
  else()
    # -a provides no cross architecture compatibility, but luckily until CUDA
    # 12.8 We have only one architecture within each family >= 9.
-    set(MIN_ARCHITECTURE_HAS_FAMILY 9999) # Effectively exclude all
-                                          # architectures
+    set(CMAKE_CUDA_MIN_ARCHITECTURE_HAS_FAMILY 9999) # Effectively exclude all
+                                                     # architectures
    set(CMAKE_CUDA_ARCHITECTURES_HAS_FAMILIES
        OFF
        PARENT_SCOPE)
  endif()
+  set(CMAKE_CUDA_MIN_ARCHITECTURE_HAS_FAMILY
+      ${CMAKE_CUDA_MIN_ARCHITECTURE_HAS_FAMILY}
+      PARENT_SCOPE)
+
  # Compatibility low bounds: Always compile kernels for these architectures. 86
  # is enabled to avoid perf regression when using 80 kernels.
  set(ARCHITECTURES_COMPATIBILITY_BASE 80 86 90 100 120)
@ -252,11 +456,11 @@ function(setup_cuda_architectures)
  set(CMAKE_CUDA_ARCHITECTURES_NORMALIZED)
  set(CMAKE_CUDA_ARCHITECTURES_FAMILIES)
  foreach(CUDA_ARCH IN LISTS CMAKE_CUDA_ARCHITECTURES_NORMALIZED_LIST)
-    if(CUDA_ARCH GREATER_EQUAL ${MIN_ARCHITECTURE_HAS_FAMILY}
+    if(CUDA_ARCH GREATER_EQUAL ${CMAKE_CUDA_MIN_ARCHITECTURE_HAS_FAMILY}
       AND NOT CUDA_ARCH IN_LIST ARCHITECTURES_NO_COMPATIBILITY)
      list(APPEND CMAKE_CUDA_ARCHITECTURES_NORMALIZED "${CUDA_ARCH}f-real")
      list(APPEND CMAKE_CUDA_ARCHITECTURES_FAMILIES "${CUDA_ARCH}f")
-    elseif(CUDA_ARCH GREATER_EQUAL ${MIN_ARCHITECTURE_HAS_ACCEL})
+    elseif(CUDA_ARCH GREATER_EQUAL ${CMAKE_CUDA_MIN_ARCHITECTURE_HAS_ACCEL})
      list(APPEND CMAKE_CUDA_ARCHITECTURES_NORMALIZED "${CUDA_ARCH}a-real")
    else()
      list(APPEND CMAKE_CUDA_ARCHITECTURES_NORMALIZED "${CUDA_ARCH}-real")
@ -271,17 +475,15 @@ function(setup_cuda_architectures)
      PARENT_SCOPE)
 endfunction()

+#[[
+Add CUDA architectures to target.
+-a suffix is added automatically for supported architectures.
+Architectures are added only if user explicitly requested support for that architecture.
+#]]
 function(add_cuda_architectures target)
-  # cmake-format: off
-  # Add CUDA architectures to target.
-  # -a suffix is added automatically for supported architectures.
-  # Architectures are added only if user explicitly requested support for that architecture.
-  # cmake-format: on
-  set(MIN_ARCHITECTURE_HAS_ACCEL 90)
-
  foreach(CUDA_ARCH IN LISTS ARGN)
    if(${CUDA_ARCH} IN_LIST CMAKE_CUDA_ARCHITECTURES_ORIG)
-      if(${CUDA_ARCH} GREATER_EQUAL ${MIN_ARCHITECTURE_HAS_ACCEL})
+      if(${CUDA_ARCH} GREATER_EQUAL ${CMAKE_CUDA_MIN_ARCHITECTURE_HAS_ACCEL})
        set(REAL_CUDA_ARCH "${CUDA_ARCH}a-real")
      else()
        set(REAL_CUDA_ARCH "${CUDA_ARCH}-real")
@ -294,18 +496,19 @@ function(add_cuda_architectures target)
  endforeach()
 endfunction()

-function(set_cuda_architectures target)
-  # cmake-format: off
-  # Set CUDA architectures for a target.
-  # -a suffix is added automatically for supported architectures.
-  # Architectures passed in may be specified with -f suffix to build family conditional version of the kernel.
-  # Non-family architectures are added only if user explicitly requested support for that architecture.
-  # Family conditional architectures are only added if user requested architectures would enable compilation for it.
-  # If user requested no architectures set on the target,
-  # the target will be compiled with `PLACEHOLDER_KERNELS` macro defined.
-  # cmake-format: on
-  set(MIN_ARCHITECTURE_HAS_ACCEL 90)
+#[[
+Set CUDA architectures for a target.

+-a suffix is added automatically for supported architectures.
+Architectures passed in may be specified with -f suffix to build family conditional version of the kernel.
+
+Non-family architectures are added only if user explicitly requested support for that architecture.
+Family conditional architectures are only added if user requested architectures would enable compilation for it.
+
+If user requested no architectures set on the target,
+the target will be compiled with `PLACEHOLDER_KERNELS` macro defined.
+#]]
+function(set_cuda_architectures target)
  set(CUDA_ARCHITECTURES "")
  foreach(CUDA_ARCH IN LISTS ARGN)
    if(${CUDA_ARCH} MATCHES "[0-9]+f")
@ -326,7 +529,7 @@ function(set_cuda_architectures target)
        endforeach()
      endif()
    elseif(${CUDA_ARCH} IN_LIST CMAKE_CUDA_ARCHITECTURES_ORIG)
-      if(${CUDA_ARCH} GREATER_EQUAL ${MIN_ARCHITECTURE_HAS_ACCEL})
+      if(${CUDA_ARCH} GREATER_EQUAL ${CMAKE_CUDA_MIN_ARCHITECTURE_HAS_ACCEL})
        list(APPEND CUDA_ARCHITECTURES "${CUDA_ARCH}a-real")
      else()
        list(APPEND CUDA_ARCHITECTURES "${CUDA_ARCH}-real")
@ -342,3 +545,153 @@ function(set_cuda_architectures target)
                                           ${CUDA_ARCHITECTURES})
  endif()
 endfunction()
+
+#[[
+Filter out source files targeting CUDA architectures not enabled.
+
+Arguments:
+  SOURCE_LIST - Name of the variable containing the list of source files to filter
+  TARGET      - Target to add compile definitions to. If the target does not exist,
+                an INTERFACE library will be created.
+  ARCHS       - List of architectures to check and potentially filter
+  IMPLICIT_FAMILY - Optional flag to enable implicit family mode
+
+For each ARCH passed in:
+
+- if IMPLICIT_FAMILY is not set:
+  - if ARCH is not suffixed by f:
+    if ARCH is not in CMAKE_CUDA_ARCHITECTURES_ORIG, source files containing "sm${ARCH}"
+    but not "sm${ARCH}f" (case insensitive) will be excluded
+    Macro "EXCLUDE_SM_${ARCH}" will be defined on TARGET
+  - if ARCH is suffixed by f, NARCH is ARCH without f suffix:
+    if ARCH is not in CMAKE_CUDA_ARCHITECTURES_FAMILIES, source files containing
+    "sm${NARCH}f" (case insensitive) will be excluded
+    Macro "EXCLUDE_SM_${NARCH}F" will be defined on TARGET
+
+- if IMPLICIT_FAMILY is set:
+  ARCH shall not suffixed by f.
+  - if ARCH >= CMAKE_CUDA_MIN_ARCHITECTURE_HAS_FAMILY:
+    if "${ARCH}f" is not in CMAKE_CUDA_ARCHITECTURES_FAMILIES,
+    source files containing "sm${ARCH}" but not "sm${ARCH}a" (case insensitive) will be excluded
+    Macro "EXCLUDE_SM_${ARCH}" (no F) will be defined on TARGET
+  - else:
+    if "${ARCH}" is not in CMAKE_CUDA_ARCHITECTURES_ORIG,
+    source files containing "sm${ARCH}" (case insensitive) will be excluded
+    Macro "EXCLUDE_SM_${ARCH}" will be defined on TARGET
+#]]
+function(filter_source_cuda_architectures)
+  set(options IMPLICIT_FAMILY)
+  set(oneValueArgs SOURCE_LIST TARGET)
+  set(multiValueArgs ARCHS)
+
+  cmake_parse_arguments(PARSE_ARGV 0 arg "${options}" "${oneValueArgs}"
+                        "${multiValueArgs}")
+  set(SOURCES "${${arg_SOURCE_LIST}}")
+
+  if(NOT TARGET ${arg_TARGET})
+    add_library(${arg_TARGET} INTERFACE)
+  endif()
+
+  # Determine if target is INTERFACE library to use correct visibility
+  get_target_property(_target_type ${arg_TARGET} TYPE)
+  if(_target_type STREQUAL "INTERFACE_LIBRARY")
+    set(_compile_def_visibility INTERFACE)
+  else()
+    set(_compile_def_visibility PUBLIC)
+  endif()
+
+  foreach(ARCH IN LISTS arg_ARCHS)
+    set(SHOULD_FILTER FALSE)
+    set(MATCH_PATTERN "")
+    set(EXCLUDE_PATTERN "")
+    set(ARCH_FOR_DEFINE "")
+
+    if(NOT arg_IMPLICIT_FAMILY)
+      # Check if ARCH ends with 'f'
+      string(REGEX MATCH "^(.+)f$" _has_f_suffix "${ARCH}")
+
+      if(_has_f_suffix)
+        # ARCH is suffixed by 'f' (e.g., "100f")
+        set(BASE_ARCH "${CMAKE_MATCH_1}")
+        if(NOT "${ARCH}" IN_LIST CMAKE_CUDA_ARCHITECTURES_FAMILIES)
+          set(SHOULD_FILTER TRUE)
+          set(ARCH_FOR_DEFINE "${BASE_ARCH}F")
+          # Match "sm${BASE_ARCH}f" - straightforward match, no exclusion
+          # pattern needed
+          set(MATCH_PATTERN ".*[Ss][Mm]_?${BASE_ARCH}f.*(cubin\.cpp|\.cu)$")
+        endif()
+      else()
+        # ARCH is NOT suffixed by 'f' (e.g., "80")
+        if(NOT "${ARCH}" IN_LIST CMAKE_CUDA_ARCHITECTURES_ORIG)
+          set(SHOULD_FILTER TRUE)
+          set(ARCH_FOR_DEFINE "${ARCH}")
+          # Match "sm${ARCH}" but NOT "sm${ARCH}f"
+          set(MATCH_PATTERN ".*[Ss][Mm]_?${ARCH}.*(cubin\.cpp|\.cu)$")
+          set(EXCLUDE_PATTERN ".*[Ss][Mm]_?${ARCH}f.*(cubin\.cpp|\.cu)$")
+        endif()
+      endif()
+    else()
+      # IMPLICIT_FAMILY is set - ARCH shall not be suffixed by 'f'
+      if(${ARCH} GREATER_EQUAL ${CMAKE_CUDA_MIN_ARCHITECTURE_HAS_FAMILY})
+        # ARCH >= CMAKE_CUDA_MIN_ARCHITECTURE_HAS_FAMILY
+        if(NOT "${ARCH}f" IN_LIST CMAKE_CUDA_ARCHITECTURES_FAMILIES)
+          set(SHOULD_FILTER TRUE)
+          set(ARCH_FOR_DEFINE "${ARCH}")
+          # Match "sm${ARCH}" but NOT "sm${ARCH}a"
+          set(MATCH_PATTERN ".*[Ss][Mm]_?${ARCH}.*(cubin\.cpp|\.cu)$")
+          set(EXCLUDE_PATTERN ".*[Ss][Mm]_?${ARCH}a.*(cubin\.cpp|\.cu)$")
+        endif()
+      else()
+        # ARCH < CMAKE_CUDA_MIN_ARCHITECTURE_HAS_FAMILY
+        if(NOT "${ARCH}" IN_LIST CMAKE_CUDA_ARCHITECTURES_ORIG)
+          set(SHOULD_FILTER TRUE)
+          set(ARCH_FOR_DEFINE "${ARCH}")
+          # Match "sm${ARCH}" - no exclusion pattern needed
+          set(MATCH_PATTERN ".*[Ss][Mm]_?${ARCH}.*(cubin\.cpp|\.cu)$")
+        endif()
+      endif()
+    endif()
+
+    if(SHOULD_FILTER)
+      # Get files matching the main pattern
+      set(SOURCES_TO_CHECK "${SOURCES}")
+      list(FILTER SOURCES_TO_CHECK INCLUDE REGEX "${MATCH_PATTERN}")
+
+      if(NOT "${EXCLUDE_PATTERN}" STREQUAL "")
+        # Find files matching the exclusion pattern (these should be kept)
+        set(SOURCES_TO_KEEP "${SOURCES_TO_CHECK}")
+        list(FILTER SOURCES_TO_KEEP INCLUDE REGEX "${EXCLUDE_PATTERN}")
+        # Remove the files we want to keep from the check list
+        if(SOURCES_TO_KEEP)
+          list(REMOVE_ITEM SOURCES_TO_CHECK ${SOURCES_TO_KEEP})
+        endif()
+      endif()
+
+      set(SOURCES_FILTERED "${SOURCES_TO_CHECK}")
+
+      list(LENGTH SOURCES_FILTERED SOURCES_FILTERED_LEN)
+      message(
+        STATUS
+          "Excluding ${SOURCES_FILTERED_LEN} cubins for SM ${ARCH} from ${CMAKE_CURRENT_SOURCE_DIR}"
+      )
+      foreach(filtered_item IN LISTS SOURCES_FILTERED)
+        message(VERBOSE "- ${filtered_item}")
+      endforeach()
+
+      # Remove filtered files from sources
+      if(SOURCES_FILTERED)
+        list(REMOVE_ITEM SOURCES ${SOURCES_FILTERED})
+      endif()
+
+      # Add compile definition to target
+      target_compile_definitions(
+        ${arg_TARGET}
+        ${_compile_def_visibility}
+        "EXCLUDE_SM_${ARCH_FOR_DEFINE}")
+    endif()
+  endforeach()
+
+  set(${arg_SOURCE_LIST}
+      "${SOURCES}"
+      PARENT_SCOPE)
+endfunction()
--- a/cpp/tensorrt_llm/common/CMakeLists.txt
+++ b/cpp/tensorrt_llm/common/CMakeLists.txt
@ -36,6 +36,7 @@ add_library(common_src OBJECT ${SRCS} ${CU_SRCS})
 add_cuda_architectures(common_src 89)
 set_property(TARGET common_src PROPERTY POSITION_INDEPENDENT_CODE ON)
 set_property(TARGET common_src PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS ON)
+target_link_libraries(common_src PUBLIC trtllm_gen_fmha_interface)

 if(ENABLE_CUBLASLT_FP4_GEMM)
  target_compile_definitions(common_src PRIVATE ENABLE_CUBLASLT_FP4_GEMM)
--- a/cpp/tensorrt_llm/kernels/CMakeLists.txt
+++ b/cpp/tensorrt_llm/kernels/CMakeLists.txt
@ -15,6 +15,20 @@
 # the License.
 #

+add_subdirectory(cutlass_kernels)
+add_subdirectory(cuteDslKernels)
+add_subdirectory(flashMLA)
+add_subdirectory(contextFusedMultiHeadAttention)
+add_subdirectory(decoderMaskedMultiheadAttention)
+add_subdirectory(selectiveScan)
+add_subdirectory(userbuffers)
+add_subdirectory(trtllmGenKernels)
+add_subdirectory(fusedLayernormKernels)
+add_subdirectory(groupRmsNormKernels)
+add_subdirectory(llama4MinLatencyKernels)
+add_subdirectory(dsv3MinLatencyKernels)
+add_subdirectory(causalConv1d)
+
 file(GLOB_RECURSE SRC_CPP *.cpp)
 file(GLOB_RECURSE SRC_CU *.cu)

@ -38,28 +52,6 @@ list(FILTER SRC_CPP EXCLUDE REGEX "userbuffers/.*")
 list(FILTER SRC_CU EXCLUDE REGEX "userbuffers/.*")
 list(FILTER SRC_CU EXCLUDE REGEX "fusedLayernormKernels/.*")

-function(filter_cuda_archs ARCH SOURCES_VAR)
-  if(NOT "${ARCH}" IN_LIST CMAKE_CUDA_ARCHITECTURES_ORIG)
-    set(FILTER_REGEX ".*[Ss][Mm]_?${ARCH}(af)?.*(cubin\.cpp|\.cu)$")
-    list(APPEND SOURCES ${${SOURCES_VAR}})
-    list(APPEND SOURCES_FILTERED ${SOURCES})
-    list(FILTER SOURCES_FILTERED INCLUDE REGEX "${FILTER_REGEX}")
-    list(LENGTH SOURCES_FILTERED SOURCES_FILTERED_LEN)
-    message(
-      STATUS
-        "Excluding ${SOURCES_FILTERED_LEN} cubins for SM ${ARCH} from ${CMAKE_CURRENT_SOURCE_DIR}"
-    )
-    foreach(filtered_item ${SOURCES_FILTERED})
-      message(VERBOSE "- ${filtered_item}")
-    endforeach()
-    list(FILTER SOURCES EXCLUDE REGEX "${FILTER_REGEX}")
-    set(${SOURCES_VAR}
-        "${SOURCES}"
-        PARENT_SCOPE)
-    add_compile_definitions("EXCLUDE_SM_${ARCH}")
-  endif()
-endfunction()
-
 if(NOT ENABLE_MULTI_DEVICE)
  list(FILTER SRC_CU EXCLUDE REGEX "customAllReduceKernels*.*cu$")
 endif()
@ -72,18 +64,5 @@ target_include_directories(
  PUBLIC
    $<TARGET_PROPERTY:${INTERNAL_CUTLASS_KERNELS_TARGET},INTERFACE_INCLUDE_DIRECTORIES>
 )
+target_link_libraries(kernels_src PUBLIC trtllm_gen_fmha_interface)
 add_cuda_architectures(kernels_src 89)
-
-add_subdirectory(cutlass_kernels)
-add_subdirectory(cuteDslKernels)
-add_subdirectory(flashMLA)
-add_subdirectory(contextFusedMultiHeadAttention)
-add_subdirectory(decoderMaskedMultiheadAttention)
-add_subdirectory(selectiveScan)
-add_subdirectory(userbuffers)
-add_subdirectory(trtllmGenKernels)
-add_subdirectory(fusedLayernormKernels)
-add_subdirectory(groupRmsNormKernels)
-add_subdirectory(llama4MinLatencyKernels)
-add_subdirectory(dsv3MinLatencyKernels)
-add_subdirectory(causalConv1d)
--- a/cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/CMakeLists.txt
+++ b/cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/CMakeLists.txt
@ -19,42 +19,56 @@ file(GLOB_RECURSE SRC_CPP *.cpp)
 file(GLOB_RECURSE SRC_CU *.cu)
 list(FILTER SRC_CU EXCLUDE REGEX "fmha_v2_cu/.*")

-filter_cuda_archs("80" SRC_CPP)
-filter_cuda_archs("86" SRC_CPP)
-filter_cuda_archs("89" SRC_CPP)
-filter_cuda_archs("90" SRC_CPP)
-filter_cuda_archs("100" SRC_CPP)
-filter_cuda_archs("120" SRC_CPP)
+add_library(context_attention_src OBJECT)

-add_library(context_attention_src OBJECT ${SRC_CPP} ${SRC_CU})
+filter_source_cuda_architectures(
+  SOURCE_LIST SRC_CPP
+  ARCHS 80 86 89 90 100 120
+  TARGET context_attention_src
+  IMPLICIT_FAMILY)
+
+target_sources(context_attention_src PRIVATE ${SRC_CPP} ${SRC_CU})
 target_compile_definitions(context_attention_src PRIVATE USE_DEMO_BERT_PARAMS=1
                                                         GENERATE_CUBIN=1)
 set_target_properties(
  context_attention_src PROPERTIES POSITION_INDEPENDENT_CODE ON
                                   CUDA_RESOLVE_DEVICE_SYMBOLS ON)
+target_link_libraries(context_attention_src PUBLIC trtllm_gen_fmha_interface)

 foreach(arch IN ITEMS 80 86 89 90 100 120)
-  if("${arch}" IN_LIST CMAKE_CUDA_ARCHITECTURES_ORIG)
-    file(GLOB arch_files "fmha_v2_cu/*_sm${arch}.cu")
-    if(arch_files)
-      set(TARGET_NAME _context_attention_kernels_${arch})
-      add_library(${TARGET_NAME} OBJECT ${arch_files})
-      target_compile_definitions(${TARGET_NAME} PRIVATE USE_DEMO_BERT_PARAMS=1
-                                                        GENERATE_CUBIN=1)
-      set_target_properties(
-        ${TARGET_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON
-                                  CUDA_RESOLVE_DEVICE_SYMBOLS ON)
-      target_include_directories(
-        ${TARGET_NAME}
-        PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../../../kernels/fmha_v2/src/
-                ${CMAKE_CURRENT_SOURCE_DIR}/../../../kernels/fmha_v2/generated/)
-      if(${arch} GREATER_EQUAL 100)
-        set_cuda_architectures(${TARGET_NAME} "${arch}f")
-      else()
-        set_cuda_architectures(${TARGET_NAME} ${arch})
-      endif()
-      target_sources(context_attention_src
-                     PUBLIC $<TARGET_OBJECTS:${TARGET_NAME}>)
+  set(ARCH_ENABLED FALSE)
+  set(TARGET_ARCH "")
+  if(${arch} GREATER_EQUAL ${CMAKE_CUDA_MIN_ARCHITECTURE_HAS_FAMILY})
+    if("${arch}f" IN_LIST CMAKE_CUDA_ARCHITECTURES_FAMILIES)
+      set(ARCH_ENABLED TRUE)
+      set(TARGET_ARCH "${arch}f")
+    endif()
+  else()
+    if("${arch}" IN_LIST CMAKE_CUDA_ARCHITECTURES_ORIG)
+      set(ARCH_ENABLED TRUE)
+      set(TARGET_ARCH "${arch}")
    endif()
  endif()
+
+  if(NOT ${ARCH_ENABLED})
+    continue()
+  endif()
+
+  file(GLOB arch_files "fmha_v2_cu/*_sm${arch}.cu")
+  if(arch_files)
+    set(TARGET_NAME _context_attention_kernels_${arch})
+    add_library(${TARGET_NAME} OBJECT ${arch_files})
+    target_compile_definitions(${TARGET_NAME} PRIVATE USE_DEMO_BERT_PARAMS=1
+                                                      GENERATE_CUBIN=1)
+    set_target_properties(
+      ${TARGET_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON
+                                CUDA_RESOLVE_DEVICE_SYMBOLS ON)
+    target_include_directories(
+      ${TARGET_NAME}
+      PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../../../kernels/fmha_v2/src/
+              ${CMAKE_CURRENT_SOURCE_DIR}/../../../kernels/fmha_v2/generated/)
+    set_cuda_architectures(${TARGET_NAME} ${TARGET_ARCH})
+    target_sources(context_attention_src
+                   PUBLIC $<TARGET_OBJECTS:${TARGET_NAME}>)
+  endif()
 endforeach()
--- a/cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/CMakeLists.txt
+++ b/cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/CMakeLists.txt
@ -27,11 +27,12 @@ set(SRC_CU_EXTRA)
 list(FILTER SRC_CPP EXCLUDE REGEX ".*nvrtcWrapper/.*")
 add_subdirectory(decoderXQAImplJIT/nvrtcWrapper)

-filter_cuda_archs("80" SRC_CPP)
-filter_cuda_archs("86" SRC_CPP)
-filter_cuda_archs("89" SRC_CPP)
-filter_cuda_archs("90" SRC_CPP)
-filter_cuda_archs("120" SRC_CPP)
+add_library(decoder_attention_src OBJECT)
+filter_source_cuda_architectures(
+  SOURCE_LIST SRC_CPP
+  ARCHS 80 86 89 90 120
+  TARGET decoder_attention_src
+  IMPLICIT_FAMILY)

 set(basic_heads 32 64 128)
 foreach(HEAD ${basic_heads})
@ -64,7 +65,7 @@ foreach(HEAD ${extra_heads})
 endforeach()

 if(NOT WIN32)
-  add_library(decoder_attention_src OBJECT ${SRC_CPP})
+  target_sources(decoder_attention_src PRIVATE ${SRC_CPP})

  # Split some sources to shared library for Linux
  add_library(${DECODER_SHARED_TARGET_0} SHARED ${SRC_CU})
@ -78,10 +79,12 @@ if(NOT WIN32)
  set_property(TARGET ${DECODER_SHARED_TARGET_1}
               PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS ON)
 else()
-  add_library(decoder_attention_src OBJECT ${SRC_CPP} ${SRC_CU} ${SRC_CU_EXTRA})
+  target_sources(decoder_attention_src PRIVATE ${SRC_CPP} ${SRC_CU}
+                                               ${SRC_CU_EXTRA})
 endif()

-target_link_libraries(decoder_attention_src PUBLIC nvrtc_wrapper_src)
+target_link_libraries(decoder_attention_src PUBLIC nvrtc_wrapper_src
+                                                   trtllm_gen_fmha_interface)
 set_property(TARGET decoder_attention_src PROPERTY POSITION_INDEPENDENT_CODE ON)
 set_property(TARGET decoder_attention_src PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS
                                                   ON)
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/CMakeLists.txt
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/CMakeLists.txt
@ -18,9 +18,13 @@
 file(GLOB_RECURSE SRC_CPP *.cpp)
 file(GLOB_RECURSE SRC_CU *.cu)

-filter_cuda_archs("100" SRC_CPP)
+add_library(trtllm_gen_batched_gemm OBJECT)
+filter_source_cuda_architectures(
+  SOURCE_LIST SRC_CPP
+  ARCHS 100 103 100f
+  TARGET trtllm_gen_batched_gemm)

-add_library(trtllm_gen_batched_gemm OBJECT ${SRC_CPP} ${SRC_CU})
+target_sources(trtllm_gen_batched_gemm PRIVATE ${SRC_CPP} ${SRC_CU})
 target_compile_definitions(trtllm_gen_batched_gemm
                           PUBLIC TLLM_GEN_EXPORT_INTERFACE TLLM_ENABLE_CUDA)

--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/BatchedGemmInterface.h
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/BatchedGemmInterface.h
@ -670,7 +670,8 @@ public:
    size_t getNumBatchedGemmConfigs() const
    {
 #ifdef TLLM_GEN_EXPORT_INTERFACE
-        return tensorrt_llm::kernels::tllmGenBatchedGemmListLen;
+        return sizeof(tensorrt_llm::kernels::tllmGenBatchedGemmList)
+            / sizeof(tensorrt_llm::kernels::tllmGenBatchedGemmList[0]);
 #else
        return 0;
 #endif
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelMetaInfo.h
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelMetaInfo.h
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelTraits.h
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelTraits.h
@ -565,13 +565,13 @@ public:

 public:
    // The MMA kind.
-    tg::MmaKind mMmaKind;
+    tg::MmaKind mMmaKind{};
    // Whether fuse Utccp into the MMA task.
-    bool mFuseUtccpWithUtcmma;
+    bool mFuseUtccpWithUtcmma{};
    // Whether use the max TMEM overlap trick.
-    bool mUseMaxTmemOverlap;
+    bool mUseMaxTmemOverlap{};
    // The number of epilogue warps.
-    int32_t mNumEpilogueWarps;
+    int32_t mNumEpilogueWarps{};
    // Helper for SMEM allocation.
    MemAllocatorHelper mSmemAllocatorHelper;
    // Helper for TMEM allocation.
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/CMakeLists.txt
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/CMakeLists.txt
@ -18,8 +18,12 @@
 file(GLOB_RECURSE SRC_CPP *.cpp)
 file(GLOB_RECURSE SRC_CU *.cu)

-filter_cuda_archs("100" SRC_CPP)
+filter_source_cuda_architectures(
+  SOURCE_LIST SRC_CPP
+  ARCHS 100 103 100f
+  TARGET trtllm_gen_fmha_interface)

 add_library(trtllm_gen_fmha OBJECT ${SRC_CPP} ${SRC_CU})
 set_property(TARGET trtllm_gen_fmha PROPERTY POSITION_INDEPENDENT_CODE ON)
 set_property(TARGET trtllm_gen_fmha PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS ON)
+target_link_libraries(trtllm_gen_fmha PUBLIC trtllm_gen_fmha_interface)
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5dbeef553388e130f1b41304228f629807dabac3c5e0c66a0c28ea9833340040
-size 630932
+oid sha256:616bdb23263627aca4ce3448e32e9b47b59439aac8f774e6957415d1be92a6e5
+size 620620
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a90e201852a6f3c911e982a9f819a6be765814c39a24b0c9f53ce4dcb4d28c33
-size 565951
+oid sha256:f5b3750593eb35e8d451dc5842cbfb9cdd1e63345f7fda202eeff31e160d3839
+size 558895
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7b3100a44dbe296bc51f7ff18efd9dd1b03649bd993a1cc54da0585208395ea0
-size 483501
+oid sha256:46463e24ad7663a32582c35cf61dc56da6b37feebefcbd5df51f2cf56b6a1a5f
+size 476471
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8ee2df4ea7203351cd5d12be09234ea39cb0a0a3d1f6a7d8f87790f5f02991cf
-size 452197
+oid sha256:05f0e9c20a152b830151736c8fe075b786a2885497644ffee53c032704fb5975
+size 445981
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:70bef448c376a3e57294f21afc296ea1e7a1f21bca158cecb995de47fcf7a9e8
-size 471975
+oid sha256:4cb6bc8d55b63523388260b43cbb7b5f32d0925103f1e04d6896dd6215129578
+size 464181
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:abc46c06eba3f60de4c29a4d06b672c0fd406f0b94e82c4828919c3ad173e094
-size 447801
+oid sha256:a96c271e3b91ff63d68a9fc187e1ff980bc80822871dd5ab803dcec488e4c18d
+size 441585
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:707f1643ecd4243edbc2552b1c14e13154f659f9653122386c7d71fd97762626
-size 626982
+oid sha256:0b5202c7e41c5478e832ebfd56ebf374574ee0b643716dafe446c4943204307b
+size 619016
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:949736949a7d8db4eaff5171d6a280fc75e7597e78d5cf53cd698bf58143599c
-size 566023
+oid sha256:80486e68e84ed71716050db50073accd0eafdad79905f857f1ba01f82836d151
+size 558203
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:620338e0232c4bc83022026e927ad4337b3438bfe5225a771d71ae821708a67d
-size 452573
+oid sha256:b3df4cae191465930320aaaaaabd302040f41d1a17609902dd866656dbfb8aea
+size 454251
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:501bf6e66db0c3a774fc7505841ccf5b02da306e43522b0a694d61953cd56864
-size 394599
+oid sha256:04a36407b42e0d90fc30a20a21fe5fbfdc70c2fe64d667af6dabe5939f6d071e
+size 387593
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f0f40f5a666044cca61e289b9c48c98977691b9e1d32a72179128ccbee15cc55
-size 436291
+oid sha256:39332cc0ebc5c940feb6821101fcc757100d5b89281c0ae3f588ad5a662dd9d3
+size 437179
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c59673a8be4362f46f2e9257ec7c4e38a4800c45b1945f9df0cb32ed0c6d43ae
-size 378317
+oid sha256:164593e1d5b88e251c405d70361ac4e74508e6ce65208ab3dd6bf7976350a103
+size 371311
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:34c1fd4fa79d1384876b968ffbf432fefdf981338a7956e296799295d1d0247c
-size 505637
+oid sha256:2b8887fed8b56ea827931b12e2452ac17901c674fa9aea0b19ff4fcf0dd951f4
+size 498607
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:055d697387425f909c5329b792c5db97a58cc847cc7755169de50f99d0ed5f73
-size 468807
+oid sha256:cd9058dd7b9dc2d549b698a84f95b868b51907d0ff517de6b1b25802f148a683
+size 462591
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:60c0b37d90c1fc5e3cb27e480d84d3b18d47cf7b1cd4e3a3af35f941100e54aa
-size 494111
+oid sha256:d36b4ac976263bf8ebce2b478021cea027568e7f93e9275e6124c5cebdede87f
+size 485527
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a3b9f3aee4f1a1d79c4876a4fec9edb0baa4c78efec3c76c04ca8bd00af76963
-size 469935
+oid sha256:121ad8d9cc1eb1b4571cdb6da67c95a8bf94277fa3bc1f959a5b2f2fac9f7cdb
+size 462931
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2f417704dbb04df29ad9019c106a1de5aa4f4415e7b7aa206600843b30645000
-size 670528
+oid sha256:15829041f11b292854fb76fed462349c720a0ef30a88c24bb6d1c33754fc8490
+size 663598
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b73aa21041034ab532412a51cf00dda4d8ffbb5d3970628dc5fc9fc0dbd39e24
-size 607571
+oid sha256:8e15720366d0d135f951818421edb4cef722909f693a80a8a5208446d9907010
+size 604167
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:dfac7eaa6285ea9b8e02a1ac7318a14a89bd04dfd9794d194c55e45583aaeb1d
-size 481715
+oid sha256:75f42cbb4afe3e9d57246039bd2cd711bf65fd95dbff64b38242f493d9b7d871
+size 473919
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b18b4478ee849f9f92ff59f68887710b0c9af43d57819821e64ab81b7c8fb1cf
-size 411875
+oid sha256:3eac4d690bc59e2024c92af584bb6e0f29455e3e4575efc37cbc96e3f3263a0e
+size 405659
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b39f08d8599260d351b6df4e142512bb75dbaa219a4a701a26863c0566d3e72a
-size 460697
+oid sha256:a3c937220ee455d17e5703bb35e82a00b27ccb12bd8861e15be014823d5f4619
+size 454481
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:59497957d8e27c087941c6bc5db6dbeabbd0e02f98b3d359d0ba347698927f8f
-size 394015
+oid sha256:8ef74aa9dabff7601631ccbcaeb5880c12bf391a1f30cd830f8bf272a218a484
+size 388587
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:55485e3210691cae102219ccc7c6dc4428a79a3115887447e8d584ac25103cf5
-size 663022
+oid sha256:9f017070eaffd5f5289a69cac49ace13da7f2cc0da94f6576b6fe38f76380bf4
+size 649580
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7fd46e418a1d9ca3563211e581bf352207e88bef95293f71a15a63f516e008e4
-size 574435
+oid sha256:b102a4b49b61c21a3b1639159476fc01b977d10464bcdd50bcfed77e5ea0ec9a
+size 565703
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8e9a80632812c1d98e05131359d8d0282a0219766f4f949257804e0da566c479
-size 593219
+oid sha256:3722c508670751474ea923f7be4bb7d4bcb4ffcdf72168b4c8abe5f9921e29b3
+size 586977
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:587b63d2d1e7fb17f0ff8ab7483e4fcba28e4944bcee169dd5d5ca2909b567c3
-size 569019
+oid sha256:dd3e414313b8aa48ca65260a855ce7e58266eb6603adf67aca52efdce74da394
+size 562803
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7584eb0eb94587d360419bf58a4df610a61b99069b044f82c9bfa8669cc1ad2f
-size 580115
+oid sha256:ec71cbae8f5d86c3e8149b4aa095023082dc2cbcf465e41254c40db711b20566
+size 572319
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a578404ca82c1fdf9be383d49d0f1e7664e935088fb33d2bc1bdb4baba166d41
-size 559885
+oid sha256:8345d6335559c4bf863948ac445bb1d6696aa38ae23cdbbc392083c329324064
+size 552091
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:01801e6a36b2ad80cc9bf4ff90586b853694271cef751f55d2c5ba5eeea401d8
-size 663760
+oid sha256:f25796409ecb2803e61f092484b88d7dc6b58539b77c48b59c69dbc11abdcc52
+size 649578
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d444b0c26d99f3dc92f58a172064f3f7c21252164e9cb127f6b5ea3cfd8f9af1
-size 608499
+oid sha256:c20f6bfb52e6f222bcbbb202a4b0fb0aab4b95e73634853119141b4ee4f63a1a
+size 596609
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a2de5904a5271e374185323f2029b11fe07edbdf22666a562cec4e9868915e57
-size 558345
+oid sha256:be293e8aee13242adb29ac239f2ff332cca328949a69b0b3815e1c8116233d9f
+size 559233
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c4bc2d2f470b216d493ecd2a9d559fef26d27d34a5c9d5a44e5b890beedbc6eb
-size 501159
+oid sha256:b585477230f60b14873d5302711770bcbeff7410398792978f69caf37ef49b56
+size 492575
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6bdec3801521171bf2fc0442ccda65a5d66e36be28ec620763974a7fd8d51c1a
-size 539695
+oid sha256:35f0da62936aecf7dbd160137f5dfc408284082c5e1751e21467ade0ef502969
+size 541371
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e29787b3c4dbe5adb56c163d3e8b89191afd58615add63acdedc8a760ca3814c
-size 483299
+oid sha256:32a0d306e94f90acd490969b36f28df7f3b00a91a25e1b97d28296d2505ec54c
+size 475503
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f0584f5e2953b9827bb31c1533f9e8a36e2998bf952f30c87630c2db48362b6a
-size 616933
+oid sha256:ea93243da8bf5dff7d14328f5c13def92bd011cc25230fd7dc8b0e5e71e04d77
+size 610691
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ce98ce92ba5132756ce4b5bd3ec350e47b80ef06e573baffdfae212ba0236d57
-size 591153
+oid sha256:ae33d408c456f0db40410e49c5575354ca2ddee7aa66fb64ed77083d5a5055f0
+size 585727
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:be20a25127519e3e4816c0722c1868f9afd8556118d0ede9668a7313b781c0be
-size 603039
+oid sha256:4586d5bb50e101f31f248608789bddcb0bbcb80abd828db5c6d531b3d527dd7d
+size 596033
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a00e5f26406e9cfcf828c84c9cb054f096bf96d238989ddffef1da097243bb49
-size 582021
+oid sha256:64eebcb394addd848d705a111224979e41e202bad6e10a968ff246294c70ea27
+size 575015
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bf56cc695ada25ad4b2707dadf77d0c98c3087e2f52ee6c89b2a9d59d438cc8b
-size 702670
+oid sha256:7235264cc8742a630cb0ef55dce3b20550fcbeea1292af972eeaac882afb3234
+size 688486
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:350653f2f4134348c19bdc0460dcff2f693023b783023a4ed6a1eacae1715dbc
-size 617191
+oid sha256:3b78df7d072083125979fb3e517a2bb564e6aec52e256091c127fcdb36f63a25
+size 610135
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:90dd023f70455b0788425890b8cd7b869b21f732e2073d6085a60fea86f8fe10
-size 586695
+oid sha256:9eb18afe39f7022496e7666125b68a6a83d1a300bbe88fba2c62584bbca5f952
+size 579691
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b77f111c0a72a1c478c4850102b729dba22c7d28ce72b950af96e6ca838e36b9
-size 512909
+oid sha256:123b2693b236e75a98978328bc5e84c4d4155e82c2de1fd693a61bb284b5197c
+size 507483
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cbc91114ea6283743c584a2c908a431d0339d1f8f3c1b62d484e2147a6c91688
-size 564889
+oid sha256:20d78ba25b66081b8ca9bc81ac72d3db8700c35c71c5cbb7981cf03ffd723872
+size 559461
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:61472723f538d24c9acd2715d33bdec5b0b398c8721f760b2d1487a18f9c56db
-size 494259
+oid sha256:f43e322151d205cb7be44aa960e9c058b1130f746a0554189323dab1c088366e
+size 488833
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4d573460b47edb6494ffaee16a7ad168fc838bc1f682276dcf883f3bf87dacd8
-size 561491
+oid sha256:bd3c554d3e609746f5c88052f610407e0ba7a77d218b98f23d1700db288d3672
+size 549849
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d04d9dc89e714b2fffd0e2c3584a5acfae5d8f09b372d73331bc6bbf8ff5e31e
-size 502949
+oid sha256:303745a8ad276cbf903057d90dc83c5d5d187b9a7fc8820896460789c04a4d55
+size 496463
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:357f912b4b3932fbd1715ff5d3630cf0f57ad1ccc8345319f66f59dfe64a244a
-size 456883
+oid sha256:322e38aed9acfc5ba5083010671a292a15866cfc7ada0a1ef04ca08d11402e0f
+size 450619
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e50308861808e757f76cb664d4834089006cb5897ca20ca7c62047681b38d58d
-size 442723
+oid sha256:ef75d9f8b4c1e439803fa420345fdd27884817117e761f3d655eea2f1d300487
+size 436507
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:07088afdd4892f3359c43cd71a036eacf99aaa7a61ff5be89cd1a14380d69f57
-size 450859
+oid sha256:f7ba48776a57bee1eaba279f88b455d71830353dd0f1c17982d38534ccc7bb79
+size 444643
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cd3ae620c3742f516c4cd35b72773bc305e98b98b1cc4e63d44d5083d30f123f
-size 436747
+oid sha256:19cac5380575aff8a39767424d38b6b463b006bbf92927f24c96da6593cdffab
+size 430531
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:10dfcdeb3f26e142320f4deaa7d3100d3b411fc45696bb21232a102b0694efc1
-size 558333
+oid sha256:254581db67314ea198da7aee258596f04f6def6d01a41fd1ef71705b04087e20
+size 549083
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bec6c7e277fa9f46c6e6650bab9c17611dc4cd1669b4d75211d233db672576b9
-size 502233
+oid sha256:18572512305e1229566e0a6eb81f88f22fa8dae442f08cfa419334b00b89e980
+size 495769
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0b6ed54b526c428b08dbe9f7e80e7673fb8a2d985ce065d9ace6730c44dbcd28
-size 441719
+oid sha256:16bd9568835544733fdff2c27b988f109b15c638ab04cadf8e672fa9831fba33
+size 442607
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fee499381c761edd8ed3c707e9e86af26b7d8112b7285d8562364445b4db15e3
-size 384533
+oid sha256:7ec6566eadcb6b1c559e140b150aa15380b69aa3a5ad6363a6f2f55e8e6a66bd
+size 378317
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:527451a03c38f8a12a2fd09c911c4ea54cc89bd7f688108ea30694f661bbb5f1
-size 424449
+oid sha256:a97ad0f83edda84db155ab182b6805ab79f1b7f1f3b9a33717aa14ffcdba02f6
+size 425337
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1e3cb444f826882d1f9fdde1d49bae9235f36fa683a9ab2b7879f82a4207d689
-size 367265
+oid sha256:d8b3b3bd9790a04368061c248476f1921afaab33635117336f4fe3f8f7621e9c
+size 361049
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:32f4a45f95b8ae84a0e3ae5a43b3b5f5c7362b1ce223819f516f4b4e73170bb7
-size 473445
+oid sha256:471d613cc4f330de90ee3adaeb3ca60311cd5f5566f6c186347954cbc27f9536
+size 468017
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3d8a3cb32086a1867cce707e2300b7dae6b3f47db799a77848534a205e6846c1
-size 458543
+oid sha256:3667df7be928d46644c990d5edf800b4267ca16d2b1af246c2ff1fcd185aa63f
+size 453117
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:aa25c1b87150f6db0800cebb0d799ec948d40e80384c5312eedb52f73a1d6618
-size 473783
+oid sha256:198938eb3da07564a649fc28ae9a1f347c4aa3b3b4d94c7df599e164c7c78153
+size 467567
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:89c8a875579b5d4ff2f53c6925a71f5f382ec4e0763dd56ec92423aab9166a94
-size 458883
+oid sha256:dba544890e95b2fad732d7de898a7bfe6bae8c9e91d0655fe7f092314a6f8956
+size 452667
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a654c16808084190287ebf34535e68dc1f4d3d1f3c082912633e372f35df0648
-size 601089
+oid sha256:817cf4afb5bd708007464a22a2ae0ce270807ba2b7e5b259d1613f9380d4f9e5
+size 594947
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1fdf46c99fb0c98db2c447a5e22075c9f3e037665a9945e61763adf2b3f5824f
-size 545605
+oid sha256:ce202d4c5578e1aad38916fb24e627f6f217c727bb3cbb201d268a59d6f1eb17
+size 538231
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:04b2dd3f1d6ce30d084c944caf3301c274424a004097efdff1a345f72083dff8
-size 470069
+oid sha256:8c14dce9f0ef34ba1bfa4dc4adc7d99b304910ca5fb785bd41b0612f9771c784
+size 462275
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6fb4cc06f718df560c8de15140cabef4e98bd1d0f35e186cf37d5e94a8256863
-size 400231
+oid sha256:28c0a61e45a428546baada7385da98a5beba03372ec0bfd4282b8edd305ba7de
+size 394803
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5b1d5f5ed80ed7d8d0c0dcf465f31ada29aeedaa85e4da693a684a5cf940d927
-size 449643
+oid sha256:dec9c27d8f9b4bd5db6f88275274c94adbd30ea2ce8220afe8fbd8b4f0dba5b6
+size 443427
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:06c57c87ed62e80008fb18fe5688f1f633c45aa0cc1a66f826c1efac62af8bdf
-size 382961
+oid sha256:d1c92e1c9d473d21e3a9bf07502e01eef0903b9183ee409ff40bfa089efa5ee1
+size 377535
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvCausalVarSeqQ128Kv128PersistentContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvCausalVarSeqQ128Kv128PersistentContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e2568162532c48e842319952122b3835e4ef014861cff4f4e37491d3703ef58a
-size 749338
+oid sha256:286fce5fedc98c25e31b2aa6f9e527364b1f634cebbcf4174a9e5c6e86d73226
+size 725312
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvCausalVarSeqQ128Kv128StaticContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvCausalVarSeqQ128Kv128StaticContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:07064296444d7ac6bfff94dcc790acd605fef68a760cbc5ead3296c880789567
-size 661516
+oid sha256:1e8d8a2baf9baa1eae8674c5f55a26346e6e004d90b0d035179de74362b79275
+size 636060
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvCausalVarSeqSkipsSoftmaxQ128Kv128PersistentContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvCausalVarSeqSkipsSoftmaxQ128Kv128PersistentContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:75be8ecca1ec7acabb23fe44d3cd08fa50d46a2600a921c8e2197ff731141607
-size 776790
+oid sha256:aaca490b5dd9c7990f603c992bb750e617d927caceb019397eb8eb4fa44194de
+size 749312
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvCausalVarSeqSkipsSoftmaxQ128Kv128StaticContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvCausalVarSeqSkipsSoftmaxQ128Kv128StaticContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b354769aa9f5659b436534f7a3cc4aa09a493858d43ab70bdc316f6b0d88d512
-size 688476
+oid sha256:645a80c48a4be552dc052a5ae103e69cf8a3b504a97d2a5b235005da1ad5261c
+size 663414
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvDenseVarSeqQ128Kv128PersistentContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvDenseVarSeqQ128Kv128PersistentContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:883f192f66396084fb325abbb832687f2f27887c0018a970e39bcae575b4aa95
-size 753232
+oid sha256:7922a772bf56cc4a42e3f8fa5ea6ea26a20092baa25e85254820ab7ebb872f0b
+size 722004
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvDenseVarSeqQ128Kv128StaticContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvDenseVarSeqQ128Kv128StaticContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b35e56a59c42ea0b6f30bb49076c0f93611fcb00cdd76dd4353b465adc0c75c9
-size 661810
+oid sha256:511da0ba6a3dda888cc95df62793f89fc236b630df0bd8a4e41921058a3ee1ec
+size 636650
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvDenseVarSeqSkipsSoftmaxQ128Kv128PersistentContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvDenseVarSeqSkipsSoftmaxQ128Kv128PersistentContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:dec9396cbb1bdd3ca6dfd831b9c957ad2fa31e0ddf2a6455b91b6aeab6c7d06b
-size 773336
+oid sha256:72c6ccb1e756e7429685f0e898e60b3b680b0893638054331b88322926109f36
+size 745610
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvDenseVarSeqSkipsSoftmaxQ128Kv128StaticContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvDenseVarSeqSkipsSoftmaxQ128Kv128StaticContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0f445f0ff16b08249c7700137d350ee6b1964ddb183bf4d5646852504a38dc61
-size 688524
+oid sha256:f59be6bd1031842c71104dd7b520c66b661b24b8763dea1d663e0cf58fcaa9c7
+size 664696
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvSlidingOrChunkedCausalVarSeqQ128Kv128PersistentContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvSlidingOrChunkedCausalVarSeqQ128Kv128PersistentContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:98265cb4cc56fe7336995f39759b8a4dbd6d35b68f9c934d746ec9304c16e488
-size 818090
+oid sha256:e91b7bac7cf917b6003e608ef39fc92f5b7ae1ee6c4e16e56cae1b572f5f4781
+size 790118
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvSlidingOrChunkedCausalVarSeqQ128Kv128StaticContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvSlidingOrChunkedCausalVarSeqQ128Kv128StaticContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:80eaf7e695a9144f3e9ae58526dba4d62e19c374e3e3f5191d4216e2276ce68b
-size 731700
+oid sha256:4c52649dc2f57b803013a52f84db1d5789258817b6c34ca17917e1c38b94a64e
+size 706540
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvSlidingOrChunkedCausalVarSeqSkipsSoftmaxQ128Kv128PersistentContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvSlidingOrChunkedCausalVarSeqSkipsSoftmaxQ128Kv128PersistentContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7919d7baf0bd29baa550b4c9331cb5b74c9184211648a2ae4309956fe4b707e1
-size 844952
+oid sha256:652213aa98b3b0141802acf20310dcf7a8f24a05e94062b4e8f82b3b1f272b38
+size 817078
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvSlidingOrChunkedCausalVarSeqSkipsSoftmaxQ128Kv128StaticContext_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvSlidingOrChunkedCausalVarSeqSkipsSoftmaxQ128Kv128StaticContext_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9de7275c32c73482f2ae5e5336051bdbc172dfe02cd35950535363775a97636a
-size 760042
+oid sha256:4d9521c6adfcd41ee2730cd1e0ed4502661955c2c2b0a5e6355d150b16492b0b
+size 735128
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqQ128Kv128StaticKeepsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqQ128Kv128StaticKeepsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5be32375b4f0595bfd691d4b01e5851737bda01669dcbd1a101db1e7fecb7f8c
-size 874036
+oid sha256:2c99f44a0f0c435436496cbe1d498f185604b701ea8fb67ad1a0e2c8ee9cc444
+size 864762
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:83558fa89e1d26bbf51093c42c3cc3af31711cb555fb496620a95deed45345a5
-size 706006
+oid sha256:09126ea87e05b001bfdcb4b410ca492a00f7f9b29cd63b021153eee0e582f8df
+size 696188
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqQ32Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqQ32Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6e9ca086d9b2d1a1996b6b948efaeb61c609f8c052b901a1dee4e3992220784f
-size 775714
+oid sha256:bffb58dea79c2d285f3c70d5950a6e6af4213674df55ab97aa3bf597bbddbf6a
+size 763528
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqQ64Kv128StaticKeepsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqQ64Kv128StaticKeepsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9849935550b57458ebbc1d8cc12f4db1b05c33af085aedd5df63b3c845ad342c
-size 773198
+oid sha256:a7d0c270d8b10533914550d7b0f251dd0ad666910bdaad35f2aa4cef257df932
+size 762986
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:42e8365c84d53022a6ef4ecfc0ac76a079fb3b4637ee7ebdb43d2f16da0f12e0
-size 660370
+oid sha256:f9e73caf1128411b4154623d89c0b18eeb08eefa1c83b0c62df124e9e035f885
+size 650306
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqSkipsSoftmaxQ128Kv128StaticKeepsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqSkipsSoftmaxQ128Kv128StaticKeepsAbForGen_cubin.cpp
@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:586cb3a876e522f5968a14b01f412fdcc5dac8d4d14d1b00721b590ac02343c0
+size 919842
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqSkipsSoftmaxQ16Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqSkipsSoftmaxQ16Kv128StaticSwapsAbForGen_cubin.cpp
@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c38ae728025373804ca8bd535fea64a34bef25b3937d0b1c09beb5a42d7d425
+size 599913
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqSkipsSoftmaxQ32Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqSkipsSoftmaxQ32Kv128StaticSwapsAbForGen_cubin.cpp
@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5833c71e1fe3984ac1ec0f9dc51c18131db4bb408760c8144e406f8e9118a053
+size 657042
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqSkipsSoftmaxQ64Kv128StaticKeepsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqSkipsSoftmaxQ64Kv128StaticKeepsAbForGen_cubin.cpp
@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:80b217cc222c46a69212e3408b94d7dd90e97e14cb7ddad07d520b295fc11989
+size 786344
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqSkipsSoftmaxQ8Kv128StaticSwapsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvCgaVarSeqSkipsSoftmaxQ8Kv128StaticSwapsAbForGen_cubin.cpp
@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df279dde0c4be2f933ca078a633ad278365853f4ce3394b132b67687f34d0482
+size 558865
--- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvVarSeqQ128Kv128StaticKeepsAbForGen_cubin.cpp
+++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32MultiCtasKvVarSeqQ128Kv128StaticKeepsAbForGen_cubin.cpp
@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f2987e51b51870e7bcfdb3e631f4cbd0f5fbe47f9a8fb8f219fb6f716d11c27a
-size 834318
+oid sha256:707c90f34cff5ff1eb2efa9fb19f1167d12aab1120bd2c38025ff8682dff53ed
+size 808220
--- a/Show More
+++ b/Show More