[Gen] Fix FT kernel when using CG

flash_attn/modules/mha.py

@@ -495,7 +495,8 @@ class MHA(nn.Module):
                         *inference_params.key_value_memory_dict[self.layer_idx],
                         inference_params.lengths_per_sample, inference_params.sequence_len_offset,
                         self.rotary_emb_dim,
-                        not self.rotary_emb.interleaved  # neox_rotary_style
+                        # neox_rotary_style
+                        (not self.rotary_emb.interleaved) if self.rotary_emb_dim > 0 else True
                     )
                     context = rearrange(context, 'b h d -> b 1 h d')
         else:
@@ -609,7 +610,8 @@ class ParallelMHA(nn.Module):
                     *inference_params.key_value_memory_dict[self.layer_idx],
                     inference_params.lengths_per_sample, inference_params.sequence_len_offset,
                     self.rotary_emb_dim,
-                    not self.rotary_emb.interleaved  # neox_rotary_style
+                    # neox_rotary_style
+                    (not self.rotary_emb.interleaved) if self.rotary_emb_dim > 0 else True
                 )
                 context = rearrange(context, 'b h d -> b 1 h d')
         if seqlen is None:

flash_attn/utils/generation.py

@@ -190,9 +190,9 @@ def seqlen_to_seqlen_type(seqlen: int) -> int:
     return 0 if seqlen < 32 else (1 if seqlen < 2048 else 2)
 
 
-def seqlen_type_to_seqlen(seqlen_type: int) -> int:
+def seqlen_type_to_max_seqlen(seqlen_type: int) -> int:
     assert seqlen_type in [0, 1, 2]
-    return 1 if seqlen_type == 0 else (32 if seqlen_type == 1 else 2048)
+    return 32 if seqlen_type == 0 else (2048 if seqlen_type == 1 else 2**32)
 
 
 @dataclass
@@ -239,9 +239,9 @@ def update_graph_cache(model, cache, batch_size, seqlen_og, max_seqlen, tensor_p
         cache.mempool = torch.cuda.graphs.graph_pool_handle()
     for s_type in range(seqlen_to_seqlen_type(seqlen_og), seqlen_to_seqlen_type(max_seqlen) + 1):
         if s_type not in cache.callables:
-            seqlen = min(max(seqlen_og, seqlen_type_to_seqlen(s_type)), max_seqlen)
+            max_seqlen_ = min(max(seqlen_og, seqlen_type_to_max_seqlen(s_type)), max_seqlen)
             cache.callables[s_type] = capture_graph(
-                model, cache.inference_params, batch_size, seqlen_og, seqlen, mempool=cache.mempool,
+                model, cache.inference_params, batch_size, max_seqlen_, mempool=cache.mempool,
                 n_warmups=n_warmups
             )
 
@@ -249,17 +249,19 @@ def update_graph_cache(model, cache, batch_size, seqlen_og, max_seqlen, tensor_p
         return cache.callables[seqlen_to_seqlen_type(seqlen)](input_ids, position_ids, seqlen)
 
     cache.run = dispatch
-    cache.inference_params.sequence_length_offset = 0  # Reset so it's not confusing
+    cache.inference_params.sequence_len_offset = 0  # Reset so it's not confusing
     return cache
 
 
-def capture_graph(model, inference_params, batch_size, seqlen_og, max_seqlen, mempool=None,
+def capture_graph(model, inference_params, batch_size, max_seqlen, mempool=None, n_warmups=2):
-                  n_warmups=2):
-    assert max_seqlen >= seqlen_og
     device = next(iter(model.parameters())).device
     input_ids = torch.full((batch_size, 1), 0, dtype=torch.long, device=device)
     position_ids = torch.full((batch_size, 1), 0, dtype=torch.long, device=device)
-    inference_params.lengths_per_sample[:] = seqlen_og
+    sequence_len_offset_og = inference_params.sequence_len_offset
+    # TD [2023-04-14]: important for correctness of the FT's attention kernel, as seqlen_cpu is
+    # used to determine the size of smem. Hence seqlen_cpu must be >= lengths_per_sample.
+    inference_params.sequence_len_offset = max_seqlen - 1
+    inference_params.lengths_per_sample[:] = max_seqlen - 1
 
     # Warmup before capture
     s = torch.cuda.Stream()
@@ -289,4 +291,5 @@ def capture_graph(model, inference_params, batch_size, seqlen_og, max_seqlen, me
         graph.replay()
         return logits
 
+    inference_params.sequence_len_offset = sequence_len_offset_og
     return run

tests/models/test_gptj.py

@@ -1,4 +1,4 @@
-import re
+import time
 
 import torch
 import pytest
@@ -9,6 +9,7 @@ from transformers.models.gptj.modeling_gptj import GPTJForCausalLM
 from flash_attn.models.gpt import GPTLMHeadModel
 from flash_attn.models.gptj import remap_state_dict_hf_gptj, gptj_config_to_gpt2_config
 from flash_attn.utils.pretrained import state_dict_from_pretrained
+from flash_attn.utils.generation import update_graph_cache
 
 
 @pytest.mark.parametrize('model_name', ["EleutherAI/gpt-j-6B"])
@@ -79,3 +80,92 @@ def test_gptj_optimized(model_name):
     print(f'HF fp16 max diff: {(logits_hf - logits_ref).abs().max().item()}')
     print(f'HF fp16 mean diff: {(logits_hf - logits_ref).abs().mean().item()}')
     assert (logits - logits_ref).abs().max().item() < 3 * (logits_hf - logits_ref).abs().max().item()
+
+
+@pytest.mark.parametrize('model_name', ["EleutherAI/gpt-j-6B"])
+def test_gptj_generation(model_name):
+    """Check that our implementation of GPT-J (with all optimizations enabled) matches the
+    HF implementation: the output of our forward pass in fp16 should be around the same as the HF
+    forward pass in fp16, when compared to the HF forward pass in fp32.
+    """
+    dtype = torch.float16
+    device = 'cuda'
+    config = gptj_config_to_gpt2_config(GPTJConfig.from_pretrained(model_name))
+    config.use_flash_attn = False  # FlashAttention doesn't support hdim 256 yet
+    config.fused_bias_fc = True
+    config.fused_mlp = True
+    config.fused_dropout_add_ln = True
+    # Only prenorm supports residual_in_fp32
+    config.residual_in_fp32 = True
+
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    eos_token_id = tokenizer.eos_token_id
+
+    torch.manual_seed(0)
+    batch_size = 1
+    seqlen = 100
+    max_length = 150
+    input_ids = torch.randint(0, config.vocab_size, (batch_size, seqlen), dtype=torch.long,
+                              device=device)
+
+    model_hf = GPTJForCausalLM.from_pretrained(model_name, torch_dtype=dtype,
+                                               device_map={"": device})
+    model_hf.eval()
+    print("HF fp16")
+    torch.cuda.synchronize()
+    start = time.time()
+    out_hf = model_hf.generate(input_ids=input_ids, max_length=max_length,
+                               return_dict_in_generate=True, output_scores=True)
+    torch.cuda.synchronize()
+    print(f'Prompt processing + decoding time: {(time.time() - start) * 1000:.0f}ms')
+    del model_hf
+
+    model_ref = GPTJForCausalLM.from_pretrained(model_name, device_map={"": device})
+    model_ref.eval()
+    with torch.no_grad():
+        logits_ref = model_ref(out_hf.sequences).logits[:, (seqlen - 1):-1]
+    del model_ref
+
+    model = GPTLMHeadModel.from_pretrained(model_name, config, device=device, dtype=dtype)
+    model.eval()
+
+    print('Without CUDA graph')
+    torch.cuda.synchronize()
+    start = time.time()
+    out = model.generate(input_ids=input_ids, max_length=max_length,
+                         eos_token_id=eos_token_id, fused_ft_kernel=True,
+                         # eos_token_id=eos_token_id, fused_ft_kernel=False,
+                         return_dict_in_generate=True, output_scores=True, timing=True,
+                         teacher_outputs=out_hf.sequences)
+    torch.cuda.synchronize()
+    print(f'Prompt processing + decoding time: {(time.time() - start) * 1000:.0f}ms')
+
+    # Capture graph outside the timing loop
+    batch_size, seqlen_og = input_ids.shape
+    model._decoding_cache = update_graph_cache(model, None, batch_size, seqlen_og, max_length)
+    print('With CUDA graph')
+    torch.cuda.synchronize()
+    start = time.time()
+    out_cg = model.generate(input_ids=input_ids, max_length=max_length,
+                            fused_ft_kernel=True, cg=True,
+                            return_dict_in_generate=True, output_scores=True, timing=True,
+                            teacher_outputs=out_hf.sequences)
+    torch.cuda.synchronize()
+    print(f'Prompt processing + decoding time: {(time.time() - start) * 1000:.0f}ms')
+
+    with torch.no_grad():
+        logits_parallel = model(out_hf.sequences).logits[:, (seqlen - 1):-1]
+    logits_hf = torch.stack(out_hf.scores, dim=1)
+    logits = torch.stack(out.scores, dim=1)
+    logits_cg = torch.stack(out_cg.scores, dim=1)
+
+    del model
+
+    hf_error = (logits_hf - logits_ref).abs().max().item()
+    assert (logits_parallel - logits_ref).abs().max().item() < 2 * hf_error
+
+    print(f'HF fp16 logits max diff: {hf_error}')
+    print(f'Logits max diff: {(logits - logits_ref).abs().max().item() }')
+    assert (logits - logits_ref).abs().max().item() < 2 * hf_error
+    print(f'Logits CG max diff: {(logits_cg - logits_ref).abs().max().item() }')
+    assert torch.equal(logits_cg, logits)