[JAX] Fix unfused GQA performance (#643)

* Fix unfused GQA perf
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Remove WAR for Check failed: reduction_kind.has_value()
Signed-off-by: Reese Wang <rewang@nvidia.com>

---------
Signed-off-by: Reese Wang <rewang@nvidia.com>

qa/L0_jax_distributed_unittest/test.sh

@@ -4,9 +4,6 @@
 
 set -xe
 
-# WAR(rewang) for the "Check failed: reduction_kind.has_value()"
-export XLA_FLAGS="${XLA_FLAGS} --xla_gpu_enable_xla_runtime_executable=true"
-
 : ${TE_PATH:=/opt/transformerengine}
 pytest -Wignore -v $TE_PATH/tests/jax/test_distributed_*
 

qa/L0_jax_unittest/test.sh

@@ -14,7 +14,5 @@ pytest -Wignore -v $TE_PATH/examples/jax/mnist
 
 # Make encoder tests to have run-to-run deterministic to have the stable CI results
 export XLA_FLAGS="${XLA_FLAGS} --xla_gpu_deterministic_ops"
-# WAR(rewang) for the "Check failed: reduction_kind.has_value()"
-export XLA_FLAGS="${XLA_FLAGS} --xla_gpu_enable_xla_runtime_executable=true"
 pytest -Wignore -v $TE_PATH/examples/jax/encoder --ignore=$TE_PATH/examples/jax/encoder/test_multiprocessing_encoder.py
 pytest -Wignore -v $TE_PATH/examples/jax/encoder/test_multiprocessing_encoder.py

transformer_engine/jax/flax/transformer.py

@@ -207,20 +207,32 @@ def core_attention(query: Array,
         key = key.astype(jnp.float32)
 
     h_q, h_kv = query.shape[-2], key.shape[-2]
-    assert (h_q % h_kv == 0) and (h_q >= h_kv)
-    group_size = h_q // h_kv
-    grouped_query = query.reshape((*query.shape[:2], h_kv, group_size, query.shape[-1]))
+    # The generated GQA kernels are slower than normal MHA kernels even when h_q == h_kv.
+    # Therefore, we have to maintain two code paths.
+    is_gqa = (h_q != h_kv)
+
+    if is_gqa:
+        assert (h_q % h_kv == 0) and (h_q >= h_kv)
+        group_size = h_q // h_kv
+        grouped_query = query.reshape((*query.shape[:2], h_kv, group_size, query.shape[-1]))
 
     if transpose_batch_sequence:
-        attn_weights = jnp.einsum('qbhgd,kbhd->bhgqk', grouped_query, key)
+        if is_gqa:
+            attn_weights = jnp.einsum('qbhgd,kbhd->bhgqk', grouped_query, key)
+        else:
+            attn_weights = jnp.einsum('qbhd,kbhd->bhqk', query, key)
     else:
-        attn_weights = jnp.einsum('bqhgd,bkhd->bhgqk', grouped_query, key)
+        if is_gqa:
+            attn_weights = jnp.einsum('bqhgd,bkhd->bhgqk', grouped_query, key)
+        else:
+            attn_weights = jnp.einsum('bqhd,bkhd->bhqk', query, key)
 
     attn_weights = checkpoint_name(attn_weights, 'logits')
 
-    b, h, g, q, k = attn_weights_with_groups_shape = attn_weights.shape
-    attn_weights_without_groups_shape = (b, h * g, q, k)
-    attn_weights = attn_weights.reshape(attn_weights_without_groups_shape)
+    if is_gqa:
+        b, h, g, q, k = attn_weights_with_groups_shape = attn_weights.shape
+        attn_weights_without_groups_shape = (b, h * g, q, k)
+        attn_weights = attn_weights.reshape(attn_weights_without_groups_shape)
 
     attn_weights = _with_sharding_constraint(attn_weights,
                                              (BATCH_AXES, HEAD_AXES, SEQLEN_AXES, SEQLEN_AXES))
@@ -237,7 +249,8 @@ def core_attention(query: Array,
     attn_weights = Softmax(softmax_type=softmax_type,
                            scale_factor=fused_scale_factor)(attn_weights, mask, bias).astype(dtype)
 
-    attn_weights = attn_weights.reshape(attn_weights_with_groups_shape)
+    if is_gqa:
+        attn_weights = attn_weights.reshape(attn_weights_with_groups_shape)
 
     if not deterministic and dropout_rate > 0.:
         keep_prob = 1.0 - dropout_rate
@@ -248,9 +261,13 @@ def core_attention(query: Array,
         attn_weights = attn_weights * multiplier
 
     if transpose_batch_sequence:
-        return jnp.einsum('bhgqk,kbhd->qbhgd', attn_weights, value).reshape(query.shape)
+        if is_gqa:
+            return jnp.einsum('bhgqk,kbhd->qbhgd', attn_weights, value).reshape(query.shape)
+        return jnp.einsum('bhqk,kbhd->qbhd', attn_weights, value)
 
-    return jnp.einsum('bhgqk,bkhd->bqhgd', attn_weights, value).reshape(query.shape)
+    if is_gqa:
+        return jnp.einsum('bhgqk,bkhd->bqhgd', attn_weights, value).reshape(query.shape)
+    return jnp.einsum('bhqk,bkhd->bqhd', attn_weights, value)
 
 
 dynamic_vector_slice_in_dim = vmap(lax.dynamic_slice_in_dim, in_axes=(None, 0, None, None))