huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-09-11 20:34:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

Upgrading the tests (TP>1 fix changes to use different kernels.)

Browse Source
This commit is contained in:
Nicolas Patry 2024-10-18 14:42:16 +02:00
parent 5ca6da15f5
commit 8673bb050d
No known key found for this signature in database
GPG Key ID: D2920555C90F704C
3 changed files with 24 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
integration-tests/models/__snapshots__/test_flash_starcoder_gptq/test_flash_starcoder_gptq.json
View File

@ -11,57 +11,57 @@
}, },
{ {
"id": 3226, "id": 3226,
"logprob": -8.9453125, "logprob": -9.0234375,
"text": " ge" "text": " ge"
}, },
{ {
"id": 21017, "id": 21017,
"logprob": -8.8515625, "logprob": -9.0859375,
"text": "ometric" "text": "ometric"
}, },
{ {
"id": 81, "id": 81,
"logprob": -0.21875, "logprob": -0.25585938,
"text": "_" "text": "_"
}, },
{ {
"id": 6009, "id": 6009,
"logprob": -1.2773438, "logprob": -2.1972656,
"text": "mean" "text": "mean"
}, },
{ {
"id": 26, "id": 26,
"logprob": -0.25195312, "logprob": -0.2998047,
"text": "(" "text": "("
}, },
{ {
"id": 62, "id": 62,
"logprob": -4.8203125, "logprob": -5.6445312,
"text": "L" "text": "L"
}, },
{ {
"id": 44, "id": 44,
"logprob": -3.7734375, "logprob": -3.0839844,
"text": ":" "text": ":"
}, },
{ {
"id": 1682, "id": 1682,
"logprob": -0.8310547, "logprob": -0.6748047,
"text": " List" "text": " List"
}, },
{ {
"id": 77, "id": 77,
"logprob": -0.22766113, "logprob": -0.3864746,
"text": "[" "text": "["
}, },
{ {
"id": 1808, "id": 1808,
"logprob": -0.46240234, "logprob": -0.9355469,
"text": "float" "text": "float"
}, },
{ {
"id": 10794, "id": 10794,
"logprob": -3.0234375, "logprob": -2.5371094,
"text": "]):" "text": "]):"
} }
], ],
@ -69,7 +69,7 @@
"tokens": [ "tokens": [
{ {
"id": 284, "id": 284,
"logprob": -0.04626465, "logprob": -1.1679688,
"special": false, "special": false,
"text": "\n " "text": "\n "
}, },

24
integration-tests/models/__snapshots__/test_flash_starcoder_gptq/test_flash_starcoder_gptq_default_params.json
View File

@ -11,57 +11,57 @@
}, },
{ {
"id": 3226, "id": 3226,
"logprob": -8.9453125, "logprob": -9.015625,
"text": " ge" "text": " ge"
}, },
{ {
"id": 21017, "id": 21017,
"logprob": -8.859375, "logprob": -9.0859375,
"text": "ometric" "text": "ometric"
}, },
{ {
"id": 81, "id": 81,
"logprob": -0.21984863, "logprob": -0.25585938,
"text": "_" "text": "_"
}, },
{ {
"id": 6009, "id": 6009,
"logprob": -1.2861328, "logprob": -2.2304688,
"text": "mean" "text": "mean"
}, },
{ {
"id": 26, "id": 26,
"logprob": -0.25219727, "logprob": -0.29760742,
"text": "(" "text": "("
}, },
{ {
"id": 62, "id": 62,
"logprob": -4.8007812, "logprob": -5.6796875,
"text": "L" "text": "L"
}, },
{ {
"id": 44, "id": 44,
"logprob": -3.7949219, "logprob": -3.0742188,
"text": ":" "text": ":"
}, },
{ {
"id": 1682, "id": 1682,
"logprob": -0.8046875, "logprob": -0.67626953,
"text": " List" "text": " List"
}, },
{ {
"id": 77, "id": 77,
"logprob": -0.22424316, "logprob": -0.38842773,
"text": "[" "text": "["
}, },
{ {
"id": 1808, "id": 1808,
"logprob": -0.46191406, "logprob": -0.9165039,
"text": "float" "text": "float"
}, },
{ {
"id": 10794, "id": 10794,
"logprob": -3.0253906, "logprob": -2.5527344,
"text": "]):" "text": "]):"
} }
], ],
@ -69,7 +69,7 @@
"tokens": [ "tokens": [
{ {
"id": 284, "id": 284,
"logprob": 0.0, "logprob": -0.048583984,
"special": false, "special": false,
"text": "\n " "text": "\n "
}, },

2
server/text_generation_server/models/flash_causal_lm.py
View File

@ -1601,8 +1601,6 @@ class FlashCausalLM(Model):
max_s = batch.max_current_length max_s = batch.max_current_length
lm_head_indices = batch.prefill_head_indices lm_head_indices = batch.prefill_head_indices
print(slots)
if cu_seqlen_prefill is None and self.max_past() is not None: if cu_seqlen_prefill is None and self.max_past() is not None:
# In decode, not prefill, we're actually overwriting the KV-cache # In decode, not prefill, we're actually overwriting the KV-cache
# in a circular buffer mode. # in a circular buffer mode.