[Excutorch][Llama] Decouple input sequence length from kv cache context length (pytorch#8047)

Pull Request resolved: pytorch#7927

Decouple max sequence length, for shape dynamism in torch.export, from sequence
length used for kv cache sizing.
ghstack-source-id: 263653316

Differential Revision: [D68448334](https://our.internmc.facebook.com/intern/diff/D68448334/)

Co-authored-by: Kimish Patel <[email protected]>

Author: pytorchbot

.ci/scripts/test_eval_llama_mmlu.sh

@@ -43,6 +43,7 @@ run_and_verify() {
 	--tasks mmlu \
 	-f 5 \
 	--max_seq_length 2048 \
+	--max_context_length 2048 \
 	--limit 5 > result.txt
 
     # Verify result.txt

.ci/scripts/test_eval_llama_wikitext.sh

@@ -41,6 +41,7 @@ run_and_verify() {
 	-kv \
 	-d fp32 \
 	--max_seq_length 2048 \
+	--max_context_length 2048 \
 	--limit 5 > result.txt
 
     # Verify result.txt

examples/models/llama/export_llama_lib.py

@@ -335,6 +335,13 @@ def build_args_parser() -> argparse.ArgumentParser:
         help="maximum length sequence to evaluate",
     )
 
+    parser.add_argument(
+        "--max_context_length",
+        type=int,
+        default=128,
+        help="maximum length of context for model to remember",
+    )
+
     parser.add_argument("-2", "--fairseq2", action="store_true")
     parser.add_argument("-v", "--verbose", action="store_true")
     parser.add_argument(
@@ -579,6 +586,7 @@ def _prepare_for_llama_export(args) -> LLMEdgeManager:
             tokenizer_path=args.tokenizer_path,
             verbose=args.verbose,
             max_seq_len=args.max_seq_length,
+            max_context_len=args.max_context_length,
             input_prune_map_path=args.input_prune_map,
             output_prune_map_path=args.output_prune_map,
             metadata_str=args.metadata,
@@ -637,6 +645,11 @@ def _validate_args(args):
     """
     TODO: Combine all the backends under --backend args
     """
+
+    if args.max_context_length < args.max_seq_length:
+        raise ValueError(
+            f"max_context_length {args.max_context_length} must be >= max_seq_len {args.max_seq_length}. max_context_length impacts kv cache size that is used to remember history, while max_seq_length refers to user prompt length. Please use --max_context_length to specify context length."
+        )
     if args.enable_dynamic_shape and (args.coreml or args.mps or args.qnn):
         raise ValueError(
             "Dynamic shape is not supported with coreml, MPS or qnn backends."
@@ -662,6 +675,7 @@ def _validate_args(args):
 
 def _export_llama(args) -> LLMEdgeManager:  # noqa: C901
     _validate_args(args)
+
     pt2e_quant_params, quantizers, quant_dtype = get_quantizer_and_quant_params(args)
 
     # export_to_edge
@@ -760,13 +774,13 @@ def _export_llama(args) -> LLMEdgeManager:  # noqa: C901
         atten = builder_exported_to_edge.model.layers[0].attention
         if args.use_qnn_sha:
             cache_shape = torch.Size(
-                (atten.max_batch_size, atten.max_seq_len, atten.head_dim)
+                (atten.max_batch_size, atten.max_context_len, atten.head_dim)
             )
         else:
             cache_shape = torch.Size(
                 (
                     atten.max_batch_size,
-                    atten.max_seq_len,
+                    atten.max_context_len,
                     atten.n_kv_heads,
                     atten.head_dim,
                 )
@@ -861,6 +875,7 @@ def _load_llama_model_metadata(
     use_sdpa_with_kv_cache: bool,
     enable_dynamic_shape: bool,
     max_seq_len: int,
+    max_context_len: int,
     n_layers: int,
     vocab_size: int,
     metadata_str: Optional[str] = None,
@@ -870,6 +885,7 @@ def _load_llama_model_metadata(
         "get_bos_id": 3 if is_fairseq2 else 1,
         "get_eos_ids": [3] if is_fairseq2 else [2],
         "get_max_seq_len": max_seq_len,
+        "get_max_context_len": max_context_len,
         "get_n_layers": n_layers,
         "get_vocab_size": vocab_size,
         "use_kv_cache": use_kv_cache,
@@ -904,6 +920,7 @@ def _load_llama_model(
     tokenizer_path: Optional[str] = None,
     verbose: bool = False,
     max_seq_len: int = 128,
+    max_context_len: int = 128,
     input_prune_map_path: Optional[str] = None,
     output_prune_map_path: Optional[str] = None,
     metadata_str: Optional[str] = None,
@@ -948,6 +965,7 @@ def _load_llama_model(
             generate_full_logits=generate_full_logits,
             fairseq2=weight_type == WeightType.FAIRSEQ2,
             max_seq_len=max_seq_len,
+            max_context_len=max_context_len,
             enable_dynamic_shape=enable_dynamic_shape,
             input_prune_map_path=input_prune_map_path,
             output_prune_map_path=output_prune_map_path,
@@ -1006,10 +1024,13 @@ def _load_llama_model(
             # pyre-fixme[6]: For 5th argument expected `ModelArgs` but got
             #  `Union[Tensor, Module]`.
             model.max_seq_len,
-            # pyre-fixme[6]: For 6th argument expected `int` but got `Union[Tensor,
+            # pyre-fixme[6]: For 6th argument expected `ModelArgs` but got
+            #  `Union[Tensor, Module]`.
+            model.max_context_len,
+            # pyre-fixme[6]: For 7th argument expected `int` but got `Union[Tensor,
             #  Module]`.
             model.n_layers,
-            # pyre-fixme[6]: For 7th argument expected `int` but got `Union[Tensor,
+            # pyre-fixme[6]: For 8th argument expected `int` but got `Union[Tensor,
             #  Module]`.
             model.vocab_size,
             metadata_str,

examples/models/llama/llama_transformer.py

@@ -91,6 +91,7 @@ class ModelArgs:
     norm_eps: float = 1e-5
     max_batch_size: int = 32
     max_seq_len: int = 2048
+    max_context_len: int = 2048
     moe: bool = False  # True to enable the MoE (Mixture of Experts)
     num_experts: int = 8  # Number of experts
     num_activated_experts: int = 2  # Number of experts to activate
@@ -163,9 +164,9 @@ def __init__(self, params: ModelArgs):
         freqs_cos, freqs_sin = self.precompute_freqs_cis(
             self.params.head_dim,
             (
-                self.params.max_seq_len  # Normal llama2.
+                self.params.max_context_len  # Normal llama2.
                 if self.params.ffn_dim_multiplier is None
-                else self.params.max_seq_len * 2  # Sharded checkpoint.
+                else self.params.max_context_len * 2  # Sharded checkpoint.
             ),
             self.params.rope_freq_base,
         )
@@ -205,7 +206,7 @@ def get_freqs(self, input_pos: Optional[torch.Tensor], seq_len: int):
                 # when KV cache is used, seqlen is most likely 1. We want to slice from the start_pos.
                 input_pos_item = input_pos[-1].item()
                 torch._check_is_size(input_pos_item)
-                torch._check(input_pos_item < self.params.max_seq_len)
+                torch._check(input_pos_item < self.params.max_context_len)
                 # pyre-ignore: Incompatible parameter type [6]: torch.narrow does expect int or Tensor
                 freqs_cos = self.freqs_cos.narrow(0, input_pos_item, seq_len)
                 # pyre-ignore: Incompatible parameter type [6]
@@ -229,15 +230,15 @@ class KVCache(nn.Module):
     def __init__(
         self,
         max_batch_size: int,
-        max_seq_length: int,
+        max_context_length: int,
         n_heads: int,
         head_dim: int,
         enable_dynamic_shape: bool,
         dtype=torch.float32,
     ):
         super().__init__()
-        self.max_seq_length = max_seq_length
-        cache_shape = (max_batch_size, n_heads, max_seq_length, head_dim)
+        self.max_context_length = max_context_length
+        cache_shape = (max_batch_size, n_heads, max_context_length, head_dim)
 
         self.max_batch_size = max_batch_size
         self.n_heads = n_heads
@@ -257,7 +258,7 @@ def update(
         if self.enable_dynamic_shape:
             start_pos = input_pos[0].item()
             torch._check_is_size(start_pos)
-            torch._check(start_pos < self.max_seq_length)
+            torch._check(start_pos < self.max_context_length)
             dim_to_slice = 2
             seq_length = k_val.size(dim_to_slice)
             # Replace the entry in the cache for this token
@@ -289,14 +290,14 @@ def __init__(
         dim: int,
         head_dim: int,
         n_rep: int,
-        max_seq_len: int,
+        max_context_len: int,
         enable_dynamic_shape: bool,
     ):
         super().__init__()
         self.dim = dim
         self.head_dim = head_dim
         self.n_rep = n_rep
-        self.max_seq_len = max_seq_len
+        self.max_context_len = max_context_len
         self.enable_dynamic_shape = enable_dynamic_shape
 
     def forward(
@@ -312,7 +313,7 @@ def forward(
         if self.enable_dynamic_shape:
             start_pos = input_pos[-1].item()
             torch._check_is_size(start_pos)
-            torch._check(start_pos < self.max_seq_len)
+            torch._check(start_pos < self.max_context_len)
             seq_length = q.size(2)
             # pyre-ignore: Incompatible parameter type [6]
             attn_mask = mask.narrow(0, start_pos, seq_length)
@@ -341,7 +342,7 @@ def __init__(self, args: ModelArgs, layer_id: int, rope: Rope):
         self.n_rep = self.n_local_heads // self.n_local_kv_heads
         self.head_dim = args.head_dim
         self.max_batch_size = args.max_batch_size
-        self.max_seq_len = args.max_seq_len
+        self.max_context_len = args.max_context_len
         self.dim = args.dim
         self.wq = nn.Linear(self.dim, self.n_heads * self.head_dim, bias=False)
         self.wk = nn.Linear(self.dim, self.n_kv_heads * self.head_dim, bias=False)
@@ -354,8 +355,8 @@ def __init__(self, args: ModelArgs, layer_id: int, rope: Rope):
 
         causal_mask = torch.tril(
             torch.ones(
-                self.max_seq_len,
-                self.max_seq_len,
+                self.max_context_len,
+                self.max_context_len,
                 dtype=torch.bool,
                 device="cpu",
             )
@@ -365,7 +366,7 @@ def __init__(self, args: ModelArgs, layer_id: int, rope: Rope):
         if self.use_kv_cache:
             self.kv_cache = KVCache(
                 args.max_batch_size,
-                args.max_seq_len,
+                args.max_context_len,
                 self.n_kv_heads,
                 self.head_dim,
                 args.enable_dynamic_shape,
@@ -374,7 +375,7 @@ def __init__(self, args: ModelArgs, layer_id: int, rope: Rope):
                 dim=self.n_local_heads * self.head_dim,
                 head_dim=self.head_dim,
                 n_rep=self.n_rep,
-                max_seq_len=self.max_seq_len,
+                max_context_len=self.max_context_len,
                 enable_dynamic_shape=args.enable_dynamic_shape,
             )
 
@@ -528,6 +529,7 @@ def __init__(self, params: ModelArgs):
         self.use_kv_cache = params.use_kv_cache
         self.generate_full_logits = params.generate_full_logits
         self.max_seq_len = params.max_seq_len
+        self.max_context_len = params.max_context_len
         self.input_prune_map = params.input_prune_map
         self.output_prune_map = params.output_prune_map
 

examples/models/llama/model.py

@@ -52,8 +52,13 @@ def __init__(self, **kwargs):
         self.input_prune_map_path = kwargs.get("input_prune_map_path", None)
         self.output_prune_map_path = kwargs.get("output_prune_map_path", None)
         self.max_seq_len = kwargs.get("max_seq_len", 128)
+        self.max_context_len = kwargs.get("max_context_len", 128)
         self.args = kwargs.get("args", None)
 
+        assert (
+            self.max_context_len >= self.max_seq_len
+        ), f"max_context_len({self.max_context_len}) must be >= max_seq_len({self.max_seq_len})"
+
         # The example is using a dummy small model with random weights for demo purpose only.
         # Follow the instruction in https://github.com/facebookresearch/llama to download the model.
         device = "cpu"
@@ -136,6 +141,7 @@ def __init__(self, **kwargs):
 
         model_args: ModelArgs = ModelArgs(
             max_seq_len=self.max_seq_len,
+            max_context_len=self.max_context_len,
             max_batch_size=1,
             use_kv_cache=self.use_kv_cache,
             use_sdpa_with_kv_cache_op=self.use_sdpa_with_kv_cache_op,
@@ -219,7 +225,7 @@ def __init__(self, **kwargs):
             window_size = int(attention_sink_params[1])
             eviction_batch_size = int(attention_sink_params[2])
 
-            assert self.args.max_seq_length == sink_size + window_size
+            assert self.args.max_context_length == sink_size + window_size
 
             self.model_ = enable_attention_sink(
                 module=self.model_,

examples/models/llama/source_transformation/attention.py

@@ -32,15 +32,15 @@ class KVCacheSHA(torch.nn.Module):
     def __init__(
         self,
         max_batch_size: int,
-        max_seq_length: int,
+        max_context_length: int,
         n_heads: int,
         head_dim: int,
         dtype=torch.float32,
     ):
         super().__init__()
 
         # a buffer per head
-        cache_shape = (max_batch_size, max_seq_length, head_dim)
+        cache_shape = (max_batch_size, max_context_length, head_dim)
         for i in range(n_heads):
             self.register_buffer(
                 f"past_k_caches_{i}",
@@ -79,7 +79,7 @@ class SDPASHA(torch.nn.Module):
     def __init__(
         self,
         max_batch_size: int,
-        max_seq_length: int,
+        max_context_length: int,
         n_heads: int,
         n_rep: int,
         head_dim: int,
@@ -90,7 +90,7 @@ def __init__(
         self.n_rep = n_rep
         self.dim = dim
         self.kv_cache = KVCacheSHA(
-            max_batch_size, max_seq_length, n_heads // n_rep, head_dim
+            max_batch_size, max_context_length, n_heads // n_rep, head_dim
         )
         self.scale_factor = math.sqrt(head_dim)
 
@@ -134,11 +134,11 @@ def __init__(self, attention_mha: nn.Module):
         self.n_rep = self.n_heads // self.n_kv_heads
         self.dim = attention_mha.dim
         self.max_batch_size = attention_mha.max_batch_size
-        self.max_seq_len = attention_mha.max_seq_len
+        self.max_context_len = attention_mha.max_context_len
         self.head_dim = attention_mha.dim // self.n_heads
         self.SDPA = SDPASHA(
             self.max_batch_size,
-            self.max_seq_len,
+            self.max_context_len,
             self.n_heads,
             self.n_rep,
             self.head_dim,
@@ -184,8 +184,8 @@ def __init__(self, attention_mha: nn.Module):
 
         causal_mask = torch.tril(
             torch.ones(
-                self.max_seq_len,
-                self.max_seq_len,
+                self.max_context_len,
+                self.max_context_len,
                 dtype=torch.bool,
                 device="cpu",
             )

examples/models/llama/source_transformation/attention_sink.py

@@ -44,8 +44,8 @@ def __init__(
             self.apply_rotary_emb_to_k = hf_apply_rotary_emb_to_k
         else:
             self.apply_rotary_emb_to_k = apply_rotary_emb_to_k
-        self.max_seq_length = window_size + sink_size
-        assert self.max_seq_length == self.params.max_seq_len
+        self.max_context_length = window_size + sink_size
+        assert self.max_context_length == self.params.max_context_len
         self.eviction_batch_size = eviction_batch_size
         self.position_shift = 0
 
@@ -54,11 +54,14 @@ def get_freqs(self, input_pos: Optional[torch.Tensor], seq_len: int):
 
         input_pos_item = input_pos.item()
         torch._check_is_size(input_pos_item)
-        if input_pos_item + self.position_shift + seq_len > self.max_seq_length:
+        if input_pos_item + self.position_shift + seq_len > self.max_context_length:
             # There are not enough spaces in the cache to store the new tokens.
             # We need to evict some old tokens and shift some recent tokens.
             num_to_evict = max(
-                input_pos_item + self.position_shift - self.max_seq_length + seq_len,
+                input_pos_item
+                + self.position_shift
+                - self.max_context_length
+                + seq_len,
                 self.eviction_batch_size,
             )
             self.position_shift -= num_to_evict  # pyre-ignore [8]
@@ -121,7 +124,7 @@ def __init__(
     ):
         super().__init__(
             max_batch_size=max_batch_size,
-            max_seq_length=window_size + sink_size,
+            max_context_length=window_size + sink_size,
             n_heads=n_heads,
             head_dim=head_dim,
             enable_dynamic_shape=enable_dynamic_shape,
@@ -148,11 +151,14 @@ def evict_tokens(self, input_pos: torch.Tensor, seq_len: int) -> int:
         """
         input_pos_item = input_pos.item()
         torch._check_is_size(input_pos_item)
-        if input_pos_item + self.position_shift + seq_len > self.max_seq_length:
+        if input_pos_item + self.position_shift + seq_len > self.max_context_length:
             # There are not enough spaces in the cache to store the new tokens.
             # We need to evict some old tokens and shift some recent tokens.
             num_to_evict = max(
-                input_pos_item + self.position_shift - self.max_seq_length + seq_len,
+                input_pos_item
+                + self.position_shift
+                - self.max_context_length
+                + seq_len,
                 self.eviction_batch_size,
             )
             num_to_keep = (