Home
last modified time | relevance | path

Searched full:max_seq_len (Results 1 – 25 of 81) sorted by relevance

1234

/aosp_15_r20/external/executorch/extension/llm/modules/test/
H A Dtest_attention.py33 self.max_seq_len = 128
54 max_seq_len=self.max_seq_len,
70 max_seq_len=self.max_seq_len,
84 max_seq_len=self.max_seq_len,
99 size=(self.max_seq_len, self.max_seq_len),
111 self.et_mha.setup_cache(1, dtype=torch.float32, max_seq_len=20)
112 self.tt_mha.setup_cache(1, dtype=torch.float32, max_seq_len=20)
145 self.et_mha.setup_cache(1, dtype=torch.float32, max_seq_len=100)
146 self.tt_mha.setup_cache(1, dtype=torch.float32, max_seq_len=100)
166 self.et_mha.setup_cache(1, dtype=torch.float32, max_seq_len=100)
[all …]
/aosp_15_r20/external/executorch/extension/llm/custom_ops/
H A Dtest_sdpa_with_kv_cache.py376 (self.n_batch, self.max_seq_len, self.n_heads_kv, self.head_dim)
379 (self.n_batch, self.max_seq_len, self.n_heads_kv, self.head_dim)
382 (self.max_seq_len, self.max_seq_len),
393 self.max_seq_len = 2048
408 max_seq_len, argument
419 self.max_seq_len = max_seq_len
495 max_seq_len = 2048
498 n_heads_kv, n_heads_q, head_dim, max_seq_len, seq_len, True
505 max_seq_len = 8
507 self._test_sdpa_common(n_heads_kv, n_heads_q, head_dim, max_seq_len, seq_len)
[all …]
/aosp_15_r20/external/pytorch/test/inductor/
H A Dtest_custom_lowering.py40 …"jagged_to_padded_dense(Tensor input, Tensor offsets, SymInt max_seq_len, Scalar pad_value) -> Ten…
43 def j2pd_meta(inp, offsets, max_seq_len, pad_value): argument
45 (offsets.shape[0] - 1, max_seq_len, inp.shape[1]),
50 def j2pd_cuda(inp, offsets, max_seq_len, pad_value): argument
52 (offsets.shape[0] - 1, max_seq_len, inp.shape[1]),
62 def j2pd_lowering(inp, offsets, max_seq_len, pad_value): argument
91 ranges=[offsets.get_size()[0] - 1, max_seq_len, inp.get_size()[1]],
136 def fn(inp, offsets, max_seq_len): argument
138 inp, offsets, max_seq_len, 60.0
143 max_seq_len = 4
[all …]
/aosp_15_r20/external/executorch/extension/llm/modules/
H A Dkv_cache.py22 max_seq_len (int): maximum sequence length model will be run with
32 max_seq_len: int,
40 max_seq_len=max_seq_len,
46 self.max_seq_len = max_seq_len
48 cache_shape = (batch_size, num_kv_heads, max_seq_len, head_dim)
50 cache_shape = (batch_size, max_seq_len, num_kv_heads, head_dim)
59 "cache_pos", torch.arange(0, self.max_seq_len), persistent=False
74 …>>> cache = KVCache(batch_size=2, max_seq_len=16, num_kv_heads=4, head_dim=32, dtype=torch.bfloat1…
108 assert (self.cache_pos[0] + seq_len) <= self.max_seq_len
139 max_seq_len=self.max_seq_len,
H A Dattention.py76 max_seq_len (int): maximum sequence length supported by the model.
104 max_seq_len: int = 4096,
133 self.max_seq_len = max_seq_len
164 self, batch_size: int, dtype: torch.dtype, max_seq_len: int
172 max_seq_len (int): maximum sequence length model will be run with.
182 max_seq_len=max_seq_len,
398 max_seq_len=child.max_seq_len,
/aosp_15_r20/external/executorch/examples/llm_pte_finetuning/
H A Drunner.py51 max_seq_len = cfg.tokenizer.max_seq_len
64 max_seq_len=max_seq_len,
87 if token_size > max_seq_len:
88 tokens = tokens[:, :max_seq_len]
90 tokens = F.pad(tokens, (0, max_seq_len - token_size), value=0)
92 if labels_size > max_seq_len:
93 labels = labels[:, :max_seq_len]
95 labels = F.pad(labels, (0, max_seq_len - labels_size), value=0)
115 max_seq_len=max_seq_len,
H A Dtraining_lib.py63 if tokenizer.max_seq_len is None:
65 "PackedDataset requires a max_seq_len to be set on the tokenizer."
67 return PackedDataset(ds, max_seq_len=tokenizer.max_seq_len, split_across_pack=False)
85 max_seq_len: int,
102 if token_size > max_seq_len:
103 tokens = tokens[:, :max_seq_len]
105 tokens = F.pad(tokens, (0, max_seq_len - token_size), value=0)
107 if labels_size > max_seq_len:
108 labels = labels[:, :max_seq_len]
110 labels = F.pad(labels, (0, max_seq_len - labels_size), value=0)
H A Dmodel_exporter.py54 max_seq_len = cfg.tokenizer.max_seq_len
62 if token_size > max_seq_len:
63 tokens = tokens[:, :max_seq_len]
65 tokens = F.pad(tokens, (0, max_seq_len - token_size), value=0)
67 if labels_size > max_seq_len:
68 labels = labels[:, :max_seq_len]
70 labels = F.pad(labels, (0, max_seq_len - labels_size), value=0)
/aosp_15_r20/external/executorch/backends/vulkan/test/op_tests/
H A Dsdpa_test.cpp137 const int max_seq_len = k_cache.size(1); in construct_attention_mask() local
220 const int max_seq_len, in test_reference_sdpa() argument
227 {batch_size, max_seq_len, num_kv_heads, head_dim}, in test_reference_sdpa()
234 for (int input_pos = start_input_pos; input_pos + sequence_len < max_seq_len; in test_reference_sdpa()
289 const int max_seq_len, in test_vulkan_sdpa() argument
294 const int init_seq_len = dynamic_seq_len ? max_seq_len : base_sequence_len; in test_vulkan_sdpa()
298 {batch_size, max_seq_len, num_kv_heads, head_dim}, in test_vulkan_sdpa()
387 input_pos + seq_len < max_seq_len; in test_vulkan_sdpa()
447 const int max_seq_len = 7; in TEST() local
456 max_seq_len, in TEST()
[all …]
/aosp_15_r20/external/executorch/examples/models/llama/
H A Dllama_transformer.py92 max_seq_len: int = 2048 variable in ModelArgs
221 max_seq_len: int,
229 self.max_seq_len = max_seq_len
250 torch._check(start_pos < self.max_seq_len)
277 self.max_seq_len = args.max_seq_len
288 self.max_seq_len,
289 self.max_seq_len,
299 args.max_seq_len,
310 max_seq_len=self.max_seq_len,
465 self.max_seq_len = params.max_seq_len
[all …]
H A Dexport_llama_lib.py556 max_seq_len=args.max_seq_length,
735 (atten.max_batch_size, atten.max_seq_len, atten.head_dim)
741 atten.max_seq_len,
820 max_seq_len: int,
829 "get_max_seq_len": max_seq_len,
863 max_seq_len: int = 128,
907 max_seq_len=max_seq_len,
944 max_seq_len=model.max_seq_len,
965 model.max_seq_len,
H A Dmain.cpp35 …output). Defaults to max_seq_len. If the number of input tokens + seq_len > max_seq_len, the outpu…
/aosp_15_r20/external/executorch/examples/models/llama/runner/
H A Dgeneration.py52 max_seq_len: int,
63 max_seq_len: max length of the output sequence, after which the output will be clipped.
69 self.max_seq_len = max_seq_len
87 max_seq_len: int,
107 while len(tokens) < max_seq_len:
163 max_seq_len=self.max_seq_len,
198 max_seq_len=self.max_seq_len,
/aosp_15_r20/external/pytorch/torch/_inductor/
H A Djagged_lowerings.py93 max_seq_len: Union[int, sympy.Expr],
102 # check=False because there may be sequences longer than max_seq_len
103 seq_idx = ops.indirect_indexing(seq, max_seq_len, check=False)
145 max_seq_len = max_lengths[0]
149 output_size = [batch_size, max_seq_len, embedding_len]
210 max_seq_len = dense_size[1]
233 max_seq_len=max_seq_len,
239 ops.index_expr(max_seq_len, offsets_dtype),
242 0.0, # jagged sequence longer than max_seq_len
/aosp_15_r20/external/executorch/examples/models/llama3_2_vision/text_decoder/
H A Dmodel.py45 self.max_seq_len = kwargs.get(
46 "max_seq_len", 8192
72 size=(self.max_seq_len, self.max_seq_len),
76 self.input_pos = torch.arange(self.max_seq_len, dtype=torch.int64)
106 max_seq_len=self.max_seq_len,
155 decoder_max_seq_len=self.max_seq_len,
187 dim_seq_len = torch.export.Dim("token_dim", min=1, max=self.max_seq_len)
/aosp_15_r20/external/executorch/examples/models/llava/
H A Dexport_llava.py87 max_seq_len=llava.text_model_args.max_seq_len,
144 max_seq_len=llava.text_model_args.max_seq_len, # This may not be right
175 token_dim_1 = Dim("token_dim_1", min=2, max=llava.text_model_args.max_seq_len)
298 …ExecuTorch with sdpa_with_kv_cache: {args.use_sdpa_with_kv_cache}, max_seq_len: {args.max_seq_len}"
302 max_seq_len=args.max_seq_len,
H A Dmodel.py41 max_seq_len: int = 768,
60 max_seq_len=max_seq_len,
261 def __init__(self, use_sdpa_with_kv_cache_op=True, max_seq_len=768): argument
263 self.max_seq_len = max_seq_len
288 self.max_seq_len,
337 dim = torch.export.Dim("token_dim", min=2, max=self.max_seq_len)
/aosp_15_r20/external/executorch/examples/models/llama3_2_vision/runner/
H A Dgeneration.py17 max_seq_len: int,
25 max_seq_len,
34 size=(max_seq_len, max_seq_len),
38 self.input_pos = torch.arange(max_seq_len)
43 max_seq_len: int,
72 while len(tokens) < max_seq_len:
/aosp_15_r20/external/executorch/examples/qualcomm/oss_scripts/llama2/model/
H A Dstatic_llama.py53 self.max_seq_len = config.max_seq_len
224 self.max_seq_len = config.max_seq_len
243 config.max_seq_len,
290 atten_mask = torch.full((self.max_batch_size, self.max_seq_len), -255.0)
299 self.max_seq_len - 1,
305 self.max_seq_len - 1,
325 "get_max_seq_len": self.max_seq_len,
/aosp_15_r20/external/executorch/backends/vulkan/runtime/graph/ops/impl/
H A DSDPA.cpp154 const int64_t max_seq_len) { in add_cache_slice_view_node() argument
158 slice_sizes.at(1) = max_seq_len; in add_cache_slice_view_node()
230 const int32_t max_seq_len = graph.size_at<int32_t>(1, k_cache); in sdpa_with_kv_cache_impl() local
244 max_seq_len); in sdpa_with_kv_cache_impl()
251 max_seq_len); in sdpa_with_kv_cache_impl()
266 cache_slice_repeated_sizes.at(1) = max_seq_len; in sdpa_with_kv_cache_impl()
295 attn_weight_full_sizes.at(2) = max_seq_len; in sdpa_with_kv_cache_impl()
296 attn_weight_full_sizes.at(3) = max_seq_len; in sdpa_with_kv_cache_impl()
/aosp_15_r20/external/executorch/backends/apple/coreml/runtime/test/
H A Dexport_stateful_model.py22 max_seq_len: int,
26 "cache", torch.zeros((max_seq_len, embedding_dim), dtype=torch.float32)
43 max_seq_len = 2
44 model = StatefulModel(embedding_dim=embedding_dim, max_seq_len=max_seq_len)
/aosp_15_r20/external/executorch/examples/models/llama/source_transformation/
H A Dattention.py137 self.max_seq_len = attention_mha.max_seq_len
141 self.max_seq_len,
187 self.max_seq_len,
188 self.max_seq_len,
/aosp_15_r20/external/pytorch/torch/testing/_internal/distributed/_tensor/
H A Dcommon_dtensor.py95 max_seq_len: int = 16 variable in ModelArgs
176 assert args.max_seq_len is not None
178 self.max_seq_len = args.max_seq_len
180 self.pos_embeddings = nn.Embedding(args.max_seq_len, args.dim)
193 assert seq_len <= self.max_seq_len
/aosp_15_r20/external/tensorflow/tensorflow/core/util/ctc/
H A Dctc_loss_calculator.h152 auto max_seq_len = seq_len(0); in CalculateLoss() local
160 max_seq_len = std::max(seq_len(b), max_seq_len); in CalculateLoss()
270 max_seq_len * num_classes * in CalculateLoss()
272 max_seq_len * 2 * (2 * num_classes + 1) * in CalculateLoss()
274 max_seq_len * in CalculateLoss()
/aosp_15_r20/external/executorch/examples/qualcomm/oss_scripts/llama3_2/
H A Dqnn_llama3_2_runner.cpp46 …output). Defaults to max_seq_len. If the number of input tokens + seq_len > max_seq_len, the outpu…

1234