promote blocksparse from prototype, make it faster (#1734)

This PR promotes block sparsity from prototype in torchao. 

Chiefly, it ports over the triton addmm blocksparse kernels from core, and makes several performance improvements to them. 

All of the numbers reported below are for an H100, with blocksize=64 and sparsity_level=0.9. The default dense baseline is 134 tok/s

1) Adds padding support to the triton kernel for dense matrices with dimension < 16, like those we run into during decoding. (214 -> 218 tok/s)
2) Changes the default [num_stages](triton-lang/triton#512) parameter from 1 to 4. This has a large effect on performance, and it seemed like the default kernel autotuning either does not modify or deems this parameter to be unimportant for some reason. (218 -> 263 tok/s). 
3) Adds an env_var, BSR_AUTOTUNE, that users can use if they want to do kernel autotuning on top of the default parameters. (263 -> 266 tok/s) This seems to matter more for bs=n compute bound workloads, where I see a reduction from 0.3855 to 0.3745s on bs=8192 prefill (roughly 3%)

So in total we are seeing a **1.985x** speedup 🚀

I've also updated the documentation to not reference prototype - planning on updating the diagram in a subsequent PR. 

### Testing
I added a new test case for the padding inputs and moved the test file out of prototype. 
```
python test/sparsity/test_sparse_api.py
```

Author: jcaip

test/prototype/test_sparse_api.py renamed to test/sparsity/test_sparse_api.py

@@ -132,8 +132,9 @@ class TestBlockSparseWeight(common_utils.TestCase):
     )
     @unittest.skipIf(not torch.cuda.is_available(), "Need CUDA available")
     @common_utils.parametrize("compile", [True, False])
-    def test_sparse(self, compile):
-        input = torch.rand((1024, 1024)).half().cuda()
+    @common_utils.parametrize("input_shape", [1, 1024])
+    def test_sparse(self, compile, input_shape):
+        input = torch.rand((input_shape, 1024)).half().cuda()
         model = (
             nn.Sequential(
                 nn.Linear(1024, 2048),
@@ -152,9 +153,7 @@ def test_sparse(self, compile):
         model[1].weight.data = create_block_sparse_tensor(M, N, 64, 0.5, torch.float16)
         dense_result = model(input)
 
-        from torchao.prototype.sparsity.superblock.blocksparse import (
-            block_sparse_weight,
-        )
+        from torchao.sparsity import block_sparse_weight
 
         sparsify_(model, block_sparse_weight(blocksize=64))
         # if compile:

torchao/_models/llama/generate.py

@@ -793,9 +793,37 @@ def ffn_or_attn_only(mod, fqn):
         from torchao.sparsity import semi_sparse_weight, sparsify_
 
         if "semi" in sparsity:
-            # TODO there is a bug here, need to fix
+            # Fixed sparsity level for 2:4
             sparsify_(model.to(device), semi_sparse_weight(), filter_fn=ffn_only)
 
+        if "bsr" in sparsity:
+            from torchao.sparsity import SupermaskLinear, block_sparse_weight
+
+            # parse "bsr-0.9-64"
+            _, sparsity_level, blocksize = sparsity.split("-")
+            sparsity_level, blocksize = float(sparsity_level), int(blocksize)
+            sparsify_(
+                model,
+                lambda x: SupermaskLinear.from_linear(
+                    x,
+                    sparsity_level=sparsity_level,
+                    blocksize=blocksize,
+                ),
+                filter_fn=ffn_only,
+            )
+            print(model)
+            sparsify_(
+                model,
+                SupermaskLinear.to_linear,
+                filter_fn=ffn_only,
+            )
+            print(model)
+
+            # Accelerate with triton bsr kernels
+            sparsify_(
+                model, block_sparse_weight(blocksize=blocksize), filter_fn=ffn_only
+            )
+
     model_size = get_model_size_in_bytes(model, ignore_embeddings=True) / 1e9
 
     if save:
@@ -810,7 +838,10 @@ def ffn_or_attn_only(mod, fqn):
         print("Compiling Model")
         global decode_one_token, prefill
         decode_one_token = torch.compile(
-            decode_one_token, mode="reduce-overhead", fullgraph=True
+            decode_one_token,
+            mode="reduce-overhead",
+            fullgraph=True,
+            dynamic=True,
         )
 
         if compile_prefill:
@@ -849,7 +880,7 @@ def ffn_or_attn_only(mod, fqn):
                 prompt = f"{B_INST} {prompt.strip()} {E_INST}"
             encoded = encode_tokens(tokenizer, prompt, bos=True, device=device)
 
-        if interactive and i >= 0:
+        if interactive and i >= 0 and prefill_size is None:
             buffer = []
             period_id = tokenizer.encode(".")[0]
             done_generating = False
@@ -919,7 +950,7 @@ def callback(x):
         device_sync(device=device)  # MKG
         t = time.perf_counter() - t0
 
-        if not interactive and demo_summarize_prompt is None:
+        if not interactive and demo_summarize_prompt is None and prefill_size is None:
             tok_list = y[0].tolist()
             # truncate text after end of string token
             tokens = (

torchao/kernel/__init__.py

@@ -1,6 +1,8 @@
+from torchao.kernel.bsr_triton_ops import bsr_dense_addmm
 from torchao.kernel.intmm import int_scaled_matmul, safe_int_mm
 
 __all__ = [
+    "bsr_dense_addmm",
     "safe_int_mm",
     "int_scaled_matmul",
 ]