Update on " [ExecuTorch][BE] Split kv cache and SDPA for better code sharing"

Summary:

Why?
We have coupled SDPA with kv cache for a while. Initially this was done
as we implemented sdpa_with_kv_cache custom op to reduce multiple copy
overheads from kv cache update. (This could have been done by having
separate custom kv cache update and custom sdpa op. Recent changes
enabled this.)
As a result of SDPA module owning kv cache, we get a) non-composable
implementation and b) harder to reuse model definition and components
from repos like tune. Output of this is that we have multiple definition
of the same model, llama, lying around in ET, TorchChat and Tune. This
diff and subsequent ones will try to move in the direction where custom
kv cache and custom sdpa become decoupled and composable, making it more
module-swap friendly with tune's model definition.

How.
Earlier PRs decoupled kv cache update from sdpa. So now
1. Decouple SDPA nn.Module from KV cache.
2. Standardize on KVCache and SDPA interface. That is KVCache and SDPA
   both operate on q, k, v in [B, # heads, seq_len, head_dim] formatted
   tensors.
3. 2 will introduce multiple tranposes when KVCache and SDPA are
   replaced by custom modules, but we will write graph pass to undo
   those.

Test Plan:
Existing tests.
Make sure perf doesnt regress

Differential Revision: [D67914054](https://our.internmc.facebook.com/intern/diff/D67914054)

[ghstack-poisoned]

Author: kimishpatel

examples/models/llama/source_transformation/quantized_kv_cache.py

@@ -207,8 +207,11 @@ def replace_kv_cache_with_quantized_kv_cache(module):
             setattr(
                 module,
                 name,
-                QuantizedKVCache.from_float(child, QuantizedCacheType.AffineAsymmetric),
-                use_custom_update_cache_op=True,
+                QuantizedKVCache.from_float(
+                    child,
+                    QuantizedCacheType.AffineAsymmetric,
+                    use_custom_update_cache_op=True,
+                ),
             )
         else:
             replace_kv_cache_with_quantized_kv_cache(child)

examples/models/llama/source_transformation/sdpa.py

@@ -9,14 +9,11 @@
 # Example script for exporting Llama2 to flatbuffer
 
 import math
-from typing import Tuple, Union
+from typing import Tuple
 
 import torch
 
 from executorch.examples.models.llama.llama_transformer import KVCache, SDPA
-from executorch.examples.models.llama.source_transformation.quantized_kv_cache import (
-    QuantizedKVCache,
-)
 
 
 class SDPACustom(torch.nn.Module):

examples/models/llama/tests/test_simple_sdpa.py

@@ -32,7 +32,6 @@ def test_simple_sdpa(self):
             enable_dynamic_shape=False,
         )
         sdpa = SDPA(
-            kv_cache=copy.deepcopy(kv_cache),
             dim=dim,
             head_dim=head_dim,
             n_rep=n_rep,
@@ -44,6 +43,11 @@ def test_simple_sdpa(self):
         key = torch.randn(1, 1, n_local_heads, head_dim)
         value = torch.randn(1, 1, n_local_heads, head_dim)
         mask = torch.randn(max_seq_length, max_seq_length)
+        query = query.transpose(1, 2)
+        key = key.transpose(1, 2)
+        value = value.transpose(1, 2)
+        key, value = kv_cache.update(input_pos, key, value)
+
         sdpa_output = sdpa(
             input_pos,
             query,
@@ -54,9 +58,7 @@ def test_simple_sdpa(self):
             mask=mask,
         )
 
-        simple_sdpa = SDPASimple(
-            kv_cache=copy.deepcopy(kv_cache), dim=dim, head_dim=head_dim, n_rep=n_rep
-        )
+        simple_sdpa = SDPASimple(dim=dim, head_dim=head_dim, n_rep=n_rep)
         simple_sdpa_output = simple_sdpa(
             input_pos, query, key, value, bsz=bsz, seqlen=seqlen, mask=mask
         )

extension/llm/export/test_export_passes.py

@@ -1,4 +1,3 @@
-import os
 import unittest
 
 import torch