Add Softmax kernel in Triton. Use softmax kernel and argmax in Llama generation.py. + Small changes

.gitignore

@@ -4,3 +4,6 @@ __pycache__/
 .pytest_cache
 **/.cache
 **/meta-llama/**/*
+
+# Virtual Environment
+venv/

README.md

@@ -0,0 +1,40 @@
+# Triton Kernels
+
+## Supported Kernels
+* flash attention
+* matmul
+* cross entropy
+
+## Contributing
+```
+python3 main.py llama_chat_completion --benchmark --ckpt_dir <model_checkpoint_path> --tokenizer_path <model_tokenizer_path>
+```
+
+
+## Getting started
+
+
+
+* Install dependencies
+
+
+
+```bash
+
+python3 -m pip install -r requirements.txt
+
+```
+
+
+
+* Download llama model
+
+
+
+```bash
+
+export HF_TOKEN=xxx
+
+huggingface-cli download meta-llama/Meta-Llama-3-8B-Instruct --local-dir $HOME/models/llama-3-8b-instruct
+
+```

kernels/__init__.py

@@ -1,6 +1,7 @@
 # from .conv import _conv, conv
 from . import blocksparse
 from .cross_entropy import _cross_entropy, cross_entropy
+from .fused_softmax import TritonSoftmax, triton_softmax
 from .flash_attention import attention
 from .matmul import _matmul, get_higher_dtype, matmul
 
@@ -12,4 +13,6 @@
     "matmul",
     "attention",
     "get_higher_dtype",
-]
+    "TritonSoftmax",
+    "triton_softmax"
+]

kernels/fused_softmax.py

@@ -0,0 +1,128 @@
+import torch 
+import triton
+import triton.language as tl
+from triton.runtime import driver
+
+
+def is_hip():
+    return triton.runtime.driver.active.get_current_target().backend == "hip"
+
+
+def is_cdna():
+    return is_hip() and triton.runtime.driver.active.get_current_target().arch in ('gfx940', 'gfx941', 'gfx942',
+                                                                                   'gfx90a', 'gfx908')
+
+
+
+
+@triton.jit
+def softmax_kernel(output_ptr, input_ptr, input_row_stride, output_row_stride, n_rows, n_cols, BLOCK_SIZE: tl.constexpr,
+                   num_stages: tl.constexpr):
+    # starting row of the program
+    row_start = tl.program_id(0)
+    row_step = tl.num_programs(0)
+    for row_idx in tl.range(row_start, n_rows, row_step, num_stages=num_stages):
+        # The stride represents how much we need to increase the pointer to advance 1 row
+        row_start_ptr = input_ptr + row_idx * input_row_stride
+        # The block size is the next power of two greater than n_cols, so we can fit each
+        # row in a single block
+        col_offsets = tl.arange(0, BLOCK_SIZE)
+        input_ptrs = row_start_ptr + col_offsets
+        # Load the row into SRAM, using a mask since BLOCK_SIZE may be > than n_cols
+        mask = col_offsets < n_cols
+        row = tl.load(input_ptrs, mask=mask, other=-float('inf'))
+        # Subtract maximum for numerical stability
+        row_minus_max = row - tl.max(row, axis=0)
+        # Note that exponentiation in Triton is fast but approximate (i.e., think __expf in CUDA)
+        numerator = tl.exp(row_minus_max)
+        denominator = tl.sum(numerator, axis=0)
+        softmax_output = numerator / denominator
+        # Write back output to DRAM
+        output_row_start_ptr = output_ptr + row_idx * output_row_stride
+        output_ptrs = output_row_start_ptr + col_offsets
+        tl.store(output_ptrs, softmax_output, mask=mask)
+
+def softmax(x):
+    device = torch.cuda.current_device()
+    properties = driver.active.utils.get_device_properties(device)
+    NUM_SM = properties["multiprocessor_count"]
+    NUM_REGS = properties["max_num_regs"]
+    SIZE_SMEM = properties["max_shared_mem"]
+    WARP_SIZE = properties["warpSize"]
+    target = triton.runtime.driver.active.get_current_target()
+    kernels = {}
+
+    n_rows, n_cols = x.shape
+
+    # The block size of each loop iteration is the smallest power of two greater than the number of columns in `x`
+    BLOCK_SIZE = triton.next_power_of_2(n_cols)
+
+    # Another trick we can use is to ask the compiler to use more threads per row by
+    # increasing the number of warps (`num_warps`) over which each row is distributed.
+    # You will see in the next tutorial how to auto-tune this value in a more natural
+    # way so you don't have to come up with manual heuristics yourself.
+    num_warps = 8
+
+    # Number of software piepling stages.
+    # num_stages = 4 if SIZE_SMEM > 200000 else 2
+    num_stages = 1
+
+    # Allocate output
+    y = torch.empty_like(x)
+
+    # pre-compile kernel to get register usage and compute thread occupancy.
+    kernel, num_programs = kernels.get(BLOCK_SIZE, (None, 0))
+    if kernel is None:
+        kernel = softmax_kernel.warmup(y, x, x.stride(0), y.stride(0), n_rows, n_cols, BLOCK_SIZE=BLOCK_SIZE,
+                                       num_stages=num_stages, num_warps=num_warps, grid=(1, ))
+        kernel._init_handles()
+        n_regs = kernel.n_regs
+        size_smem = kernel.metadata.shared
+        if is_hip():
+            # NUM_REGS represents the number of regular purpose registers. On CDNA architectures this is half of all registers available.
+            # However, this is not always the case. In most cases all registers can be used as regular purpose registers.
+            # ISA SECTION (3.6.4 for CDNA3)
+            # VGPRs are allocated out of two pools: regular VGPRs and accumulation VGPRs. Accumulation VGPRs are used
+            # with matrix VALU instructions, and can also be loaded directly from memory. A wave may have up to 512 total
+            # VGPRs, 256 of each type. When a wave has fewer than 512 total VGPRs, the number of each type is flexible - it is
+            # not required to be equal numbers of both types.
+            if is_cdna():
+                NUM_GPRS = NUM_REGS * 2
+
+            # MAX_NUM_THREADS represents maximum number of resident threads per multi-processor.
+            # When we divide this number with WARP_SIZE we get maximum number of waves that can
+            # execute on a CU (multi-processor)  in parallel.
+            MAX_NUM_THREADS = properties["max_threads_per_sm"]
+            max_num_waves = MAX_NUM_THREADS // WARP_SIZE
+            occupancy = min(NUM_GPRS // WARP_SIZE // n_regs, max_num_waves) // num_warps
+        else:
+            occupancy = NUM_REGS // (n_regs * WARP_SIZE * num_warps)
+        occupancy = min(occupancy, SIZE_SMEM // size_smem)
+        num_programs = NUM_SM * occupancy
+        kernels[BLOCK_SIZE] = (kernel, num_programs)
+
+    num_programs = min(num_programs, n_rows)
+
+    # Create a number of persistent programs.
+    kernel[(num_programs, 1, 1)](
+        y,
+        x,
+        x.stride(0),
+        y.stride(0),
+        n_rows,
+        n_cols,
+    )
+    return y
+
+
+
+class TritonSoftmax(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x):
+        return softmax(x)
+
+    # @staticmethod
+    # def backward(ctx, grad_output):
+    #     return grad_output, grad_output
+
+triton_softmax = TritonSoftmax.apply

main.py

@@ -7,6 +7,7 @@
 from models.llama import llama_example_chat_completion, llama_example_text_completion
 from benchmarking import Profiler, compare_benchmarks
 import pprint
+import torch.distributed as dist
 
 
 def main(operation: str, profile=False, benchmark=False, **kwargs):
@@ -65,6 +66,9 @@ def main(operation: str, profile=False, benchmark=False, **kwargs):
         print(output)
         print("\n==================================\n")
 
+    if dist.is_initialized():
+        dist.destroy_process_group()
+
 
 def runner(operation: str, kwargs):
     if operation == "llama_chat_completion":

models/llama/llama/generation.py

@@ -19,7 +19,7 @@
 from .math_ops import MathOps
 from .tokenizer import ChatFormat, Dialog, Message, Tokenizer
 from benchmarking import Profiler
-
+from kernels.fused_softmax import triton_softmax
 
 class CompletionPrediction(TypedDict, total=False):
     generation: str
@@ -193,10 +193,17 @@ def generate(
         for cur_pos in range(min_prompt_len, total_len):
             logits = self.model.forward(tokens[:, prev_pos:cur_pos], prev_pos)
             if temperature > 0:
-                probs = self.Math.softmax(logits[:, -1] / temperature, dim=-1)
+                if self.use_triton: 
+                    probs = triton_softmax(logits[:,-1])
+                else: 
+                    probs = self.Math.softmax(logits[:, -1] / temperature, dim=-1)
+
                 next_token = sample_top_p(probs, top_p)
             else:
-                next_token = self.Math.argmax(logits[:, -1], dim=-1)
+                if self.use_triton: 
+                    next_token = self.triton.language.argmax(logits[:, -1], axis=-1)
+                else:
+                    next_token = self.Math.argmax(logits[:, -1], dim=-1)
 
             next_token = next_token.reshape(-1)
             # only replace token if prompt has already been generated

test/README.md

@@ -0,0 +1,3 @@
+# Unit Tests for Kernels
+
+Invoke by calling ```pytest test/```

test/test_softmax.py

@@ -0,0 +1,19 @@
+import torch
+import torch.nn.functional as F
+import pytest
+from kernels.fused_softmax import triton_softmax
+
+@pytest.mark.parametrize("input_size", [(1024, 1024), (512, 512), (2048, 512)])
+def test_softmax_equivalence(input_size):
+    # Create random input tensor of specified size
+    x = torch.randn(*input_size).cuda()
+
+    # Compute softmax using PyTorch
+    pytorch_softmax = F.softmax(x, dim=-1)
+
+    # Compute softmax using Triton
+    triton_output = triton_softmax(x)
+
+    # Assert that both outputs are approximately equal
+    assert torch.allclose(pytorch_softmax, triton_output, atol=1e-5), \
+        f"Triton softmax output doesn't match PyTorch softmax for input size {input_size}"