feat: add liger kernel with fused cross entropy loss

plugins/framework/src/fms_acceleration/framework_plugin.py

@@ -206,7 +206,7 @@ def _check_config_and_maybe_check_values(
                 t = list(t.keys())[0]  # otherwise take the first value
 
             if t not in values:
-                if default is None:
+                if t is not None or default is None:
                     raise AccelerationPluginConfigError(
                         f"{self.__class__.__name__}: Value at '{key}' was '{t}'. "
                         f"Not found in expected set '{values}'."

plugins/fused-ops-and-kernels/.isort.cfg

@@ -10,4 +10,5 @@ known_firstparty=
 known_localfolder=tuning
 
 # skip code imported from unsloth
-skip_glob=**/unsloth*/**
+skip_glob=**/unsloth*/**,
+          **/liger*/**

plugins/fused-ops-and-kernels/.pylintrc

@@ -53,7 +53,9 @@ ignore=CVS,protobufs
 # format. Because '\\' represents the directory delimiter on Windows systems,
 # it can't be used as an escape character.
 # NOTE: do not lint code imported from unsloth
-ignore-paths=.*fused_ops/unsloth_lora.*,.*kernels/unsloth*
+ignore-paths=.*fused_ops/unsloth_lora.*,
+             .*fused_ops/liger_ce.*,
+             .*kernels/unsloth*,
 
 # Files or directories matching the regular expression patterns are skipped.
 # The regex matches against base names, not paths. The default value ignores

plugins/fused-ops-and-kernels/README.md

@@ -79,10 +79,23 @@ It is realtively easy by following an existing template, in what follows we use
     )
     ```
 
+### Running Liger Kernel Benchmarks
+
+Using the [scenarios-liger.yaml](../../scripts/benchmarks/scenarios-liger.yaml), this will run full fine tuning, lora peft, autoGPTQ lora peft, and bits-and-bytes lora peft with the triton kernels (Fast RMS, RoPE, CrossEnt) as a base and then run with the liger kernel for LigerFusedLinearCrossEntropy as well as Fast RMS, RoPE to compare results. It only runs against mistral and llama models.
+
+The benchmarks were ran separately for each `num_gpu` entry; they can be run together in a single command, but this is more efficient.
+
+```sh
+tox -e run-benches -- 1 "4 8 16 32" benchmark_outputs_1 scenarios-liger.yaml none
+tox -e run-benches -- 2 "8 16 32 64" benchmark_outputs_2 scenarios-liger.yaml none
+tox -e run-benches -- 4 "16 32 64 128" benchmark_outputs_3 scenarios-liger.yaml none
+```
+
+
 ## Known Issues
 
 - MixedPrecision `--fp16` or `--bf16` should be used with `fast_lora`.
 - `fast_lora` has issues with FSDP V1 with the `peft` style of FSDP wrapping. 
     * This is because the adapter's forward functions are bypassed in the fused ops.
     * For AutoGPTQ/QLoRA this is addressed by distributing the adapters using DDP so they will be unsharded in time for the fused ops.
-- `fast_rope_embeddings` does not work with position_ids. Currently `position_ids` are ignored and could give wrong results.
+- `fast_rope_embeddings` does not work with `postion_ids`, it seems like HF has depracated passing these ids into the rope embedding methods.

plugins/fused-ops-and-kernels/configs/fast_kernels.yaml

@@ -22,4 +22,4 @@ training:
     fast_rms_layernorm: True
 
     # fast RoPE embedding triton kernels
-    fast_rope_embeddings: True
+    fast_rope_embeddings: True

plugins/fused-ops-and-kernels/configs/fast_kernels_liger.yaml

@@ -0,0 +1,25 @@
+training:
+
+  fused_ops_and_kernels: 
+
+    # if under training stanza, then putting
+    # base_layer and fused_lora will be a misnomer
+    # - this should be in peft.quantized
+    # However, if it is specified, it will still 
+    # be read. This is useful in use cases where
+    # the yaml is system generated and not shown
+    # to a user.
+
+    # activate various unsloth optimizations
+    # there are two versions of the plugin
+    # - the FastKernel version supports individual kernels
+    # - the FastQuantized version is all-or-nothing
+
+    # fast loss triton kernels
+    fast_loss: fused_ce_liger
+
+    # fast rms norm triton kernels
+    fast_rms_layernorm: True
+
+    # fast RoPE embedding triton kernels
+    fast_rope_embeddings: True

plugins/fused-ops-and-kernels/configs/fast_quantized_peft_liger.yaml

@@ -0,0 +1,30 @@
+# PEFT-related acceleration
+peft:
+
+  # quantization-releated acceleration
+  # e.g., kernels for quantized base weights
+  quantization: 
+
+    fused_ops_and_kernels: 
+
+      # load unsloth optimizations for these 4bit base layer weights.
+      # currently only support "auto_gptq" and "bitsandbytes"
+      base_layer: auto_gptq
+
+      # activate various unsloth optimizations
+      # there are two versions of the plugin
+      # - the FastKernel version supports individual kernels
+      # - the FastQuantized version is all-or-nothing
+
+
+      # fused kernels for lora linear layers
+      fused_lora: True
+
+      # fast loss triton kernels
+      fast_loss: fused_ce_liger
+
+      # fast rms norm triton kernels
+      fast_rsm_layernorm: True
+
+      # fast RoPE embedding triton kernels
+      fast_rope_embeddings: True

plugins/fused-ops-and-kernels/pyproject.toml

@@ -29,3 +29,11 @@ only-include = ["src/fms_acceleration_foak"]
 
 [tool.hatch.build.targets.wheel.sources]
 "src" = ""
+
+[tool.black]
+force-exclude = '''
+/(
+.*unsloth.*
+| .*liger.*
+)/
+'''

plugins/fused-ops-and-kernels/src/fms_acceleration_foak/framework_plugin_fast_kernels.py

@@ -23,8 +23,8 @@
 import torch
 
 # Local
-from .utils import lora_adapters_switch_ddp_from_fsdp
 from .models.utils import filter_mp_rules
+from .utils import lora_adapters_switch_ddp_from_fsdp
 
 
 # consider rewriting register_foak_model_patch_rules into something
@@ -73,7 +73,10 @@ def register_foak_model_patch_rules(
 # maybe this we should define envvars
 FILTER_MAP = {
     "fused_lora": {"qkvo", "mlp"},
-    "fast_loss": "cross-ent",
+    "fast_loss": {
+        True: "cross-ent",
+        "fused_ce_liger": "fused-lce",
+    },
     "fast_rms_layernorm": "rms",
     "fast_rope_embeddings": "rope",
 }
@@ -109,19 +112,19 @@ def __init__(self, configurations: Dict[str, Dict]):
             key="base_layer", values=["auto_gptq", "bitsandbytes"], default="auto_gptq"
         )
         self.configurations["fused_lora"] = self._check_config_and_maybe_check_values(
-            key="fused_lora", values=[False, True], default=True
+            key="fused_lora", values=[False, True], default=False
         )
         self.configurations["fast_loss"] = self._check_config_and_maybe_check_values(
-            key="fast_loss", values=[False, True], default=True
+            key="fast_loss", values=[False, True, "fused_ce_liger"], default=False
         )
         self.configurations["fast_rms_layernorm"] = (
             self._check_config_and_maybe_check_values(
-                key="fast_rms_layernorm", values=[False, True], default=True
+                key="fast_rms_layernorm", values=[False, True], default=False
             )
         )
         self.configurations["fast_rope_embeddings"] = (
             self._check_config_and_maybe_check_values(
-                key="fast_rope_embeddings", values=[False, True], default=True
+                key="fast_rope_embeddings", values=[False, True], default=False
             )
         )
 
@@ -162,6 +165,8 @@ def augmentation(
 
             if k in FILTER_MAP and k not in omitted:
                 ts = FILTER_MAP[k]
+                if isinstance(ts, dict) and v in ts:
+                    ts = ts[v]
                 if isinstance(ts, str):
                     ts = {ts}
 

plugins/fused-ops-and-kernels/src/fms_acceleration_foak/fused_ops/liger_ce/__init__.py

@@ -0,0 +1,27 @@
+# Copyright 2024 Byron Hsu & Linkedin team. All rights reserved.
+#
+# BSD 2-CLAUSE LICENSE
+# Copyright 2024 LinkedIn Corporation
+# All Rights Reserved.
+# Redistribution and use in source and binary forms, with or
+# without modification, are permitted provided that the following
+# conditions are met:
+# 1. Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+# 2. Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following
+# disclaimer in the documentation and/or other materials provided
+# with the distribution.
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+from .fused_linear_cross_entropy_loss import lce_forward