huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-06-19 15:52:08 +00:00
Code Issues Packages Projects Releases Wiki Activity

format

Browse Source
This commit is contained in:
OlivierDehaene 2022-12-15 16:50:11 +01:00
parent 8420ee8fa2
commit c90afea3de
5 changed files with 57 additions and 46 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
router/src/server.rs
View File

@ -64,14 +64,14 @@ async fn health(state: Extension<ServerState>) -> Result<(), (StatusCode, Json<E
/// Generate method /// Generate method
#[instrument( #[instrument(
skip(state), skip(state),
fields( fields(
total_time, total_time,
validation_time, validation_time,
queue_time, queue_time,
inference_time, inference_time,
time_per_token time_per_token
) )
)] )]
async fn generate( async fn generate(
state: Extension<ServerState>, state: Extension<ServerState>,
@ -123,7 +123,7 @@ async fn generate(
tokens, tokens,
}) })
} }
false => None false => None,
}; };
// Timings // Timings
@ -164,7 +164,6 @@ async fn generate(
tracing::Span::current().record("time_per_token", format!("{:?}", time_per_token)); tracing::Span::current().record("time_per_token", format!("{:?}", time_per_token));
tracing::info!("Output: {}", response.output_text); tracing::info!("Output: {}", response.output_text);
// Send response // Send response
let response = vec![GeneratedText { let response = vec![GeneratedText {
generated_text: response.output_text, generated_text: response.output_text,

16
server/tests/models/test_bloom.py
View File

@ -128,7 +128,9 @@ def test_causal_lm_generate_token_completion(default_bloom, default_bloom_batch)
assert next_batch is None assert next_batch is None
assert len(generated_texts) == 1 assert len(generated_texts) == 1
assert generated_texts[0].output_text == "TestTestTestTestTestTestTestTestTestTestTest" assert (
generated_texts[0].output_text == "TestTestTestTestTestTestTestTestTestTestTest"
)
assert generated_texts[0].request == default_bloom_batch.requests[0] assert generated_texts[0].request == default_bloom_batch.requests[0]
assert ( assert (
generated_texts[0].generated_tokens generated_texts[0].generated_tokens
@ -170,7 +172,9 @@ def test_causal_lm_generate_token_completion_multi(
assert next_batch is None assert next_batch is None
assert len(generated_texts) == 1 assert len(generated_texts) == 1
assert generated_texts[0].output_text == "TestTestTestTestTestTestTestTestTestTestTest" assert (
generated_texts[0].output_text == "TestTestTestTestTestTestTestTestTestTestTest"
)
assert generated_texts[0].request == default_multi_requests_bloom_batch.requests[0] assert generated_texts[0].request == default_multi_requests_bloom_batch.requests[0]
assert ( assert (
generated_texts[0].generated_tokens generated_texts[0].generated_tokens
@ -259,7 +263,9 @@ def test_batch_concatenate(
assert next_batch is not None assert next_batch is not None
assert len(generated_texts) == 1 assert len(generated_texts) == 1
assert generated_texts[0].output_text == "TestTestTestTestTestTestTestTestTestTestTest" assert (
generated_texts[0].output_text == "TestTestTestTestTestTestTestTestTestTestTest"
)
assert generated_texts[0].request == default_bloom_batch.requests[0] assert generated_texts[0].request == default_bloom_batch.requests[0]
assert ( assert (
generated_texts[0].generated_tokens generated_texts[0].generated_tokens
@ -279,7 +285,9 @@ def test_batch_concatenate(
assert next_batch is None assert next_batch is None
assert len(generated_texts) == 1 assert len(generated_texts) == 1
assert generated_texts[0].output_text == "TestTestTestTestTestTestTestTestTestTestTest" assert (
generated_texts[0].output_text == "TestTestTestTestTestTestTestTestTestTestTest"
)
assert generated_texts[0].request == default_multi_requests_bloom_batch.requests[0] assert generated_texts[0].request == default_multi_requests_bloom_batch.requests[0]
assert ( assert (
generated_texts[0].generated_tokens generated_texts[0].generated_tokens

22
server/text_generation/models/causal_lm.py
View File

@ -148,8 +148,8 @@ class CausalLMBatch:
# We need to slice the attention mask to remove padding from previous steps # We need to slice the attention mask to remove padding from previous steps
attention_mask[ attention_mask[
start_index:end_index, -batch.max_sequence_length: start_index:end_index, -batch.max_sequence_length :
] = batch.attention_mask[:, -batch.max_sequence_length:] ] = batch.attention_mask[:, -batch.max_sequence_length :]
for j, past in enumerate(batch.past_key_values): for j, past in enumerate(batch.past_key_values):
past_keys, past_values = past past_keys, past_values = past
@ -199,20 +199,20 @@ class CausalLMBatch:
past_key_values[j][0][ past_key_values[j][0][
start_index:end_index, start_index:end_index,
:, :,
-(batch.max_sequence_length - 1):, -(batch.max_sequence_length - 1) :,
:, :,
] = past_keys[:, :, -(batch.max_sequence_length - 1):, :] ] = past_keys[:, :, -(batch.max_sequence_length - 1) :, :]
else: else:
past_key_values[j][0][ past_key_values[j][0][
start_index:end_index, start_index:end_index,
:, :,
:, :,
-(batch.max_sequence_length - 1):, -(batch.max_sequence_length - 1) :,
] = past_keys[:, :, :, -(batch.max_sequence_length - 1):] ] = past_keys[:, :, :, -(batch.max_sequence_length - 1) :]
past_key_values[j][1][ past_key_values[j][1][
start_index:end_index, :, -(batch.max_sequence_length - 1):, : start_index:end_index, :, -(batch.max_sequence_length - 1) :, :
] = past_values[:, :, -(batch.max_sequence_length - 1):, :] ] = past_values[:, :, -(batch.max_sequence_length - 1) :, :]
start_index += batch.size start_index += batch.size
@ -355,7 +355,9 @@ class CausalLM(Model):
token_ids = all_input_ids[-new_input_length:] token_ids = all_input_ids[-new_input_length:]
tokens = self.tokenizer.batch_decode(token_ids) tokens = self.tokenizer.batch_decode(token_ids)
# Add NaN for the first prompt token # Add NaN for the first prompt token
logprobs = [float('nan')] + all_logprobs[-new_input_length:].squeeze(1).tolist() logprobs = [float("nan")] + all_logprobs[-new_input_length:].squeeze(
1
).tolist()
# Add to the list of finished generations with the original request # Add to the list of finished generations with the original request
generated_texts.append( generated_texts.append(
@ -366,7 +368,7 @@ class CausalLM(Model):
tokens=tokens, tokens=tokens,
token_ids=token_ids.squeeze(1).tolist(), token_ids=token_ids.squeeze(1).tolist(),
logprobs=logprobs, logprobs=logprobs,
reason=reason reason=reason,
) )
) )
# add to the next batch # add to the next batch

6
server/text_generation/models/seq2seq_lm.py
View File

@ -450,7 +450,9 @@ class Seq2SeqLM(Model):
tokens = self.tokenizer.batch_decode(token_ids) tokens = self.tokenizer.batch_decode(token_ids)
print(tokens) print(tokens)
# Add NaN for the bos token # Add NaN for the bos token
logprobs = [float('nan')] + decoder_logprobs[-new_decoder_input_length:].tolist() logprobs = [float("nan")] + decoder_logprobs[
-new_decoder_input_length:
].tolist()
# Add to the list of finished generations with the original request # Add to the list of finished generations with the original request
generated_texts.append( generated_texts.append(
GeneratedText( GeneratedText(
@ -460,7 +462,7 @@ class Seq2SeqLM(Model):
tokens=tokens, tokens=tokens,
token_ids=token_ids.tolist(), token_ids=token_ids.tolist(),
logprobs=logprobs, logprobs=logprobs,
reason=reason reason=reason,
) )
) )
# add to the next batch # add to the next batch