[C/PyTorch/Jax] Add support for more bias shapes (#677)

* added support for arbitrary bias shapes for fused_attn Signed-off-by: Alp Dener <[email protected]> * Fix linting Signed-off-by: Alp Dener <[email protected]> * Add b1ss/bhss/11ss bias shapes when not requiring dBias Signed-off-by: Charlene Yang <[email protected]> * fix lint Signed-off-by: Charlene Yang <[email protected]> * add bias_b/h to plan cache Signed-off-by: Charlene Yang <[email protected]> * fixed compile errors after PR653 merge Signed-off-by: Alp Dener <[email protected]> * updated JAX unittests for new bias shapes Signed-off-by: Alp Dener <[email protected]> * fixed mismatched mask type checking Signed-off-by: Alp Dener <[email protected]> * corrected skip condition Signed-off-by: Alp Dener <[email protected]> * fix selection logic for A100s Signed-off-by: Charlene Yang <[email protected]> * corrected skip checks for bias shapes Signed-off-by: Alp Dener <[email protected]> * resolved test issues but neginf with float16 is still problematic with JAX Signed-off-by: Alp Dener <[email protected]> * new bias shapes passing TE JAX CI for seqlen <= 512, seq_q == seq_kv and h_q == h_kv conditions Signed-off-by: Alp Dener <[email protected]> * TE/JAX fused attn tests for new bias shapes passing with neg_inf=-2**27 for Bfloat16 and -2**15 for Float16 Signed-off-by: Alp Dener <[email protected]> * code style fixes and test parameter ID cleanup Signed-off-by: Alp Dener <[email protected]> * fixed incorrect skip condition for backward fused attn test Signed-off-by: Alp Dener <[email protected]> --------- Signed-off-by: Alp Dener <[email protected]> Signed-off-by: Charlene Yang <[email protected]> Co-authored-by: Alp Dener <[email protected]>
NVIDIA · Feb 28, 2024 · b8eea8a · b8eea8a
1 parent 0404095
commit b8eea8a
Show file tree

Hide file tree

Showing 9 changed files with 508 additions and 266 deletions.
diff --git a/tests/jax/test_fused_attn.py b/tests/jax/test_fused_attn.py
diff --git a/tests/pytorch/fused_attn/test_fused_attn.py b/tests/pytorch/fused_attn/test_fused_attn.py
@@ -85,6 +85,7 @@ def __init__(
         attn_bias_type: str,
         alibi_type: str = "none",
         num_layers: int = 1,
+        bias_shape: str = "1hss",
     ):
         self.batch_size = batch_size
         self.num_heads = num_heads
@@ -100,6 +101,7 @@ def __init__(
         self.alibi_type  = alibi_type
         self.attn_type  = "self" if (max_seqlen_q == max_seqlen_kv) else "cross"
         self.num_layers = num_layers
+        self.bias_shape = bias_shape
 
 def _is_fused_attention_supported(
     config: ModelConfig,
@@ -379,6 +381,31 @@ def test_dpa_bias(dtype, model_configs, model):
     """Test DotProductAttention module with different bias types"""
     test_dot_product_attention(dtype, model_configs, model, False, True, None, False)
 
+model_configs_bias_shapes = {
+    #     test:             b,  h, hg,   d,   sq,  skv,   p,
+    "bias_1_0": ModelConfig(4, 16, 16,  64,  128,  128, 0.0,
+    #        mask,                     bias,       bias_shape,
+        "no_mask",        "post_scale_bias", bias_shape='11ss'),
+    "bias_1_1": ModelConfig(2, 16, 16,  64,  128,  128, 0.0,
+        "no_mask",        "post_scale_bias", bias_shape='1hss'),
+    "bias_1_2": ModelConfig(4, 24, 24, 128, 2048, 2048, 0.0,
+        "no_mask",         "post_scale_bias", bias_shape='b1ss'),
+    "bias_1_3": ModelConfig(2, 24, 24, 128, 2048, 2048, 0.0,
+        "no_mask",         "post_scale_bias", bias_shape='bhss'),
+    "bias_1_4": ModelConfig(4, 24, 24, 128, 2048, 2048, 0.0,
+        "causal",                   "alibi", bias_shape='1hss', alibi_type='custom'),
+    "bias_1_5": ModelConfig(2, 24, 24, 128, 2048, 2048, 0.0,
+        "causal",                   "alibi", bias_shape='bhss', alibi_type='custom'),
+}
+
+@pytest.mark.skipif(_cudnn_version() < (8,9,1), reason="cuDNN 8.9.1+ is required.")
+@pytest.mark.parametrize("dtype", param_types_lean)
+@pytest.mark.parametrize("model_configs", [model_configs_bias_shapes])
+@pytest.mark.parametrize("model", model_configs_bias_shapes.keys())
+def test_dpa_bias_shapes(dtype, model_configs, model):
+    """Test DotProductAttention module with different bias types and shapes"""
+    test_dot_product_attention(dtype, model_configs, model, False, True, None, False)
+
 model_configs_swa = {
     #     test:             b,  h, hg,   d,   sq,  skv,   p,             mask,             bias
     "swa_1_0": ModelConfig(4, 16, 16,  64,  128,  128, 0.0,        "no_mask",          "no_bias"),
@@ -510,10 +537,13 @@ def _run_dot_product_attention(
         window_size, attention_mask = None, None
 
     alibi_slopes = None
-    if config.attn_bias_type == "alibi":
-        if config.alibi_type == "custom":
+    if config.attn_bias_type == "alibi" and config.alibi_type == "custom":
+        if config.bias_shape == "1hss":
             alibi_slopes = torch.randn(
                 config.num_heads).abs().to(dtype=torch.float32, device="cuda")
+        if config.bias_shape == "bhss":
+            alibi_slopes = torch.randn(
+                config.batch_size, config.num_heads).abs().to(dtype=torch.float32, device="cuda")
 
     # Create input tensors
     dim_to_num = {
@@ -527,6 +557,7 @@ def _run_dot_product_attention(
         'tg' : cu_seqlens_kv[-1],
         '3'  : 3,
         '2'  : 2,
+        '1'  : 1,
         }
     inp = []
     for i,layout in enumerate(qkv_layout.split('_')):
@@ -566,8 +597,12 @@ def _run_dot_product_attention(
     if config.attn_bias_type in ['no_bias', 'alibi']:
         bias = None
     if config.attn_bias_type == 'post_scale_bias':
-        bias = torch.randn(1, config.num_heads, config.max_seqlen_q, config.max_seqlen_kv,
-                dtype=dtype, device="cuda")
+        shape = '_'.join(config.bias_shape)
+        shape = shape.replace('_s_s', '_sq_skv')
+        tensor_shape = [dim_to_num[j] for j in shape.split('_')]
+        bias = torch.randn(tensor_shape, dtype=dtype, device="cuda")
+        if config.bias_shape != '1hss':
+            bias.requires_grad = False
 
     # Create RNG
     _DUMMY_CUDA_RNG_STATE_TRACKER = CudaRNGStatesTracker()

diff --git a/transformer_engine/common/fused_attn/fused_attn_f16_arbitrary_seqlen.cu b/transformer_engine/common/fused_attn/fused_attn_f16_arbitrary_seqlen.cu
@@ -72,6 +72,7 @@ void fused_attn_arbitrary_seqlen_fwd_impl(
         FADescriptor_v1 descriptor{b,                   h,
                                    hg,                  s_q,
                                    s_kv,                d,
+                                   bias_b,              bias_h,
                                    scaling_factor,      is_training,
                                    dropout_probability, layout,
                                    bias_type,           mask_type,
@@ -316,6 +317,7 @@ void fused_attn_arbitrary_seqlen_bwd_impl(
         FADescriptor_v1 descriptor{b,                   h,
                                    hg,                  s_q,
                                    s_kv,                d,
+                                   bias_b,              bias_h,
                                    scaling_factor,      true,
                                    dropout_probability, layout,
                                    bias_type,           mask_type,
@@ -426,7 +428,12 @@ void fused_attn_arbitrary_seqlen_bwd_impl(
                                 .set_dim({bias_b, bias_h, s_q, s_kv})
                                 .set_stride({bias_h * s_q * s_kv, s_q * s_kv, s_kv, 1}));
                 sdpa_backward_options.set_bias(bias);
-                sdpa_backward_options.set_dbias(dBias);
+                // shapes [1, 1, s, s], [b, 1, s, s], [b, h, s, s]
+                // are not supported for dbias calculation but they are
+                // supported for forward bias calculation
+                if ((bias_b == 1) && (bias_h == h)) {
+                  sdpa_backward_options.set_dbias(dBias);
+                }
             }
 
             if (is_padding) {
@@ -541,7 +548,11 @@ void fused_attn_arbitrary_seqlen_bwd_impl(
 
         if (is_bias) {
             variant_pack[bias] = devPtrBias;
-            variant_pack[dBias] = devPtrdBias;
+            if ((bias_b == 1) && (bias_h == h)) {
+              variant_pack[dBias] = devPtrdBias;
+            } else {
+              variant_pack[dBias] = nullptr;
+            }
         }
 
         if (is_padding) {

diff --git a/transformer_engine/common/fused_attn/utils.h b/transformer_engine/common/fused_attn/utils.h
@@ -103,6 +103,8 @@ struct FADescriptor_v1 {
   std::int64_t s_q;
   std::int64_t s_kv;
   std::int64_t d;
+  std::int64_t bias_b;
+  std::int64_t bias_h;
   float attnScale;
   bool isTraining;
   float dropoutProbability;
@@ -112,11 +114,12 @@ struct FADescriptor_v1 {
   cudnn_frontend::DataType_t tensor_type;
 
   bool operator<(const FADescriptor_v1 &rhs) const {
-    return std::tie(b, h, hg, s_q, s_kv, d,
+    return std::tie(b, h, hg, s_q, s_kv, d, bias_b, bias_h,
                     attnScale, isTraining, dropoutProbability,
                     layout, mask_type, bias_type, tensor_type)
                     < std::tie(
                       rhs.b, rhs.h, rhs.hg, rhs.s_q, rhs.s_kv, rhs.d,
+                      rhs.bias_b, rhs.bias_h,
                       rhs.attnScale, rhs.isTraining,
                       rhs.dropoutProbability, rhs.layout,
                       rhs.mask_type, rhs.bias_type,

diff --git a/transformer_engine/common/include/transformer_engine/fused_attn.h b/transformer_engine/common/include/transformer_engine/fused_attn.h
@@ -44,7 +44,7 @@ enum NVTE_QKV_Layout {
 };
 
 /*! \enum NVTE_QKV_Layout_Group
- *  \brief QKV layout groups 
+ *  \brief QKV layout groups
  */
 enum NVTE_QKV_Layout_Group {
     /*! 3HD QKV layouts, i.e. BS3HD, SB3HD, T3HD */