Note
Go to the end to download the full example code.
Steel method forward to guess the dynamic shapes (with Tiny-LLM)¶
Inputs are always dynamic with LLMs that is why dynamic shapes
needs to be specified when a LLM is exported with:func:torch.export.export.
Most of the examples on HuggingFace use method
transformers.GenerationMixin.generate() but we only want to
export the model and its method forward.
That example shows to guess the inputs of this method even though the model
is executed through meth generate.
We focus on the model arnir0/Tiny-LLM. To avoid downloading any weights, we write a function creating a random model based on the same architecture.
Steel the forward method¶
The first step is to guess the dummy inputs. Let’s use the true model for that. We use the dummy example from the model page.
import copy
import pprint
import torch
import transformers
from onnx_diagnostic import doc
from onnx_diagnostic.helpers import string_type
from onnx_diagnostic.helpers.torch_test_helper import steel_forward
from onnx_diagnostic.torch_models.llms import get_tiny_llm
MODEL_NAME = "arnir0/Tiny-LLM"
tokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_NAME)
model = transformers.AutoModelForCausalLM.from_pretrained(MODEL_NAME)
We rewrite the forward method to print the cache dimension.
def _forward_(*args, _f=None, **kwargs):
assert _f is not None
if not hasattr(torch.compiler, "is_exporting") or not torch.compiler.is_exporting():
# torch.compiler.is_exporting requires torch>=2.7
print("<-", string_type((args, kwargs), with_shape=True, with_min_max=True))
res = _f(*args, **kwargs)
if not hasattr(torch.compiler, "is_exporting") or not torch.compiler.is_exporting():
print("->", string_type(res, with_shape=True, with_min_max=True))
return res
keep_model_forward = model.forward
model.forward = lambda *args, _f=keep_model_forward, **kwargs: _forward_(
*args, _f=_f, **kwargs
)
Let’s run the model.
prompt = "Continue: it rains..."
inputs = tokenizer.encode(prompt, return_tensors="pt")
outputs = model.generate(
inputs, max_length=50, temperature=1, top_k=50, top_p=0.95, do_sample=True
)
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
print("-- prompt", prompt)
print("-- answer", generated_text)
<- ((),dict(cache_position:T7s8[0,7:A3.5],past_key_values:DynamicCache[serialized](#2[#0[],#0[]]),input_ids:T7s1x8[1,29901:A6305.375],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x8x32000[-15.516718864440918,15.75765609741211:A-3.381915190983544],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x8x96[-5.490959167480469,6.226877689361572:A-0.11321351693110653]],#1[T1s1x1x8x96[-0.6787744760513306,0.49568021297454834:A0.007227749521139988]]]))
<- ((),dict(cache_position:T7s1[8,8:A8.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x8x96[-5.490959167480469,6.226877689361572:A-0.11321351693110653]],#1[T1s1x1x8x96[-0.6787744760513306,0.49568021297454834:A0.007227749521139988]]]),input_ids:T7s1x1[13,13:A13.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-10.432564735412598,8.368535995483398:A-4.234468644971028],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x9x96[-5.509540557861328,6.348220348358154:A-0.12195695057461206]],#1[T1s1x1x9x96[-0.6787744760513306,0.7704185843467712:A0.009565710057611594]]]))
<- ((),dict(cache_position:T7s1[9,9:A9.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x9x96[-5.509540557861328,6.348220348358154:A-0.12195695057461206]],#1[T1s1x1x9x96[-0.6787744760513306,0.7704185843467712:A0.009565710057611594]]]),input_ids:T7s1x1[3112,3112:A3112.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-15.801546096801758,9.950498580932617:A-6.4666844126652],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x10x96[-5.509540557861328,6.348220348358154:A-0.13962116193664162]],#1[T1s1x1x10x96[-0.6787744760513306,0.7704185843467712:A0.0072120334078590535]]]))
<- ((),dict(cache_position:T7s1[10,10:A10.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x10x96[-5.509540557861328,6.348220348358154:A-0.13962116193664162]],#1[T1s1x1x10x96[-0.6787744760513306,0.7704185843467712:A0.0072120334078590535]]]),input_ids:T7s1x1[471,471:A471.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-19.286500930786133,6.964541435241699:A-8.171564785795635],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x11x96[-5.509540557861328,6.348220348358154:A-0.14506798955622027]],#1[T1s1x1x11x96[-0.6787744760513306,0.7704185843467712:A0.00736443510189047]]]))
<- ((),dict(cache_position:T7s1[11,11:A11.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x11x96[-5.509540557861328,6.348220348358154:A-0.14506798955622027]],#1[T1s1x1x11x96[-0.6787744760513306,0.7704185843467712:A0.00736443510189047]]]),input_ids:T7s1x1[278,278:A278.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-18.826133728027344,5.930154323577881:A-7.345287751398515],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x12x96[-5.509540557861328,6.348220348358154:A-0.1527417454230494]],#1[T1s1x1x12x96[-0.6787744760513306,0.7704185843467712:A0.007889281518436696]]]))
<- ((),dict(cache_position:T7s1[12,12:A12.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x12x96[-5.509540557861328,6.348220348358154:A-0.1527417454230494]],#1[T1s1x1x12x96[-0.6787744760513306,0.7704185843467712:A0.007889281518436696]]]),input_ids:T7s1x1[1833,1833:A1833.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-20.171358108520508,9.203018188476562:A-7.78908254127251],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x13x96[-5.509540557861328,6.348220348358154:A-0.14717222428841453]],#1[T1s1x1x13x96[-0.6787744760513306,0.7704185843467712:A0.007065091680522948]]]))
<- ((),dict(cache_position:T7s1[13,13:A13.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x13x96[-5.509540557861328,6.348220348358154:A-0.14717222428841453]],#1[T1s1x1x13x96[-0.6787744760513306,0.7704185843467712:A0.007065091680522948]]]),input_ids:T7s1x1[310,310:A310.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-15.129871368408203,8.837909698486328:A-6.361443364663515],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x14x96[-5.509540557861328,6.348220348358154:A-0.13477496815409853]],#1[T1s1x1x14x96[-0.6787744760513306,0.7704185843467712:A0.007072468134310839]]]))
<- ((),dict(cache_position:T7s1[14,14:A14.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x14x96[-5.509540557861328,6.348220348358154:A-0.13477496815409853]],#1[T1s1x1x14x96[-0.6787744760513306,0.7704185843467712:A0.007072468134310839]]]),input_ids:T7s1x1[278,278:A278.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-16.48991584777832,3.820629596710205:A-8.514456785946619],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x15x96[-5.509540557861328,6.348220348358154:A-0.12910320374770284]],#1[T1s1x1x15x96[-0.6787744760513306,0.7704185843467712:A0.0075118097320531286]]]))
<- ((),dict(cache_position:T7s1[15,15:A15.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x15x96[-5.509540557861328,6.348220348358154:A-0.12910320374770284]],#1[T1s1x1x15x96[-0.6787744760513306,0.7704185843467712:A0.0075118097320531286]]]),input_ids:T7s1x1[937,937:A937.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-17.814903259277344,5.775240898132324:A-9.154742403260897],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x16x96[-5.509540557861328,6.348220348358154:A-0.1324728711708758]],#1[T1s1x1x16x96[-0.6787744760513306,0.7704185843467712:A0.007435579083211981]]]))
<- ((),dict(cache_position:T7s1[16,16:A16.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x16x96[-5.509540557861328,6.348220348358154:A-0.1324728711708758]],#1[T1s1x1x16x96[-0.6787744760513306,0.7704185843467712:A0.007435579083211981]]]),input_ids:T7s1x1[2211,2211:A2211.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-20.981536865234375,5.375881195068359:A-8.965695807254873],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x17x96[-5.509540557861328,6.348220348358154:A-0.13291970647375873]],#1[T1s1x1x17x96[-0.6787744760513306,0.7704185843467712:A0.006298902270979549]]]))
<- ((),dict(cache_position:T7s1[17,17:A17.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x17x96[-5.509540557861328,6.348220348358154:A-0.13291970647375873]],#1[T1s1x1x17x96[-0.6787744760513306,0.7704185843467712:A0.006298902270979549]]]),input_ids:T7s1x1[3064,3064:A3064.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-16.166728973388672,9.487751007080078:A-7.495344128720462],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x18x96[-5.509540557861328,6.348220348358154:A-0.12991640532864376]],#1[T1s1x1x18x96[-0.6787744760513306,0.7704185843467712:A0.007811267364314955]]]))
<- ((),dict(cache_position:T7s1[18,18:A18.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x18x96[-5.509540557861328,6.348220348358154:A-0.12991640532864376]],#1[T1s1x1x18x96[-0.6787744760513306,0.7704185843467712:A0.007811267364314955]]]),input_ids:T7s1x1[297,297:A297.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-17.668142318725586,6.795965194702148:A-8.826972996983677],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x19x96[-5.509540557861328,7.693107604980469:A-0.12875871387433635]],#1[T1s1x1x19x96[-0.6787744760513306,0.7704185843467712:A0.00787170972308447]]]))
<- ((),dict(cache_position:T7s1[19,19:A19.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x19x96[-5.509540557861328,7.693107604980469:A-0.12875871387433635]],#1[T1s1x1x19x96[-0.6787744760513306,0.7704185843467712:A0.00787170972308447]]]),input_ids:T7s1x1[278,278:A278.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-19.28282928466797,3.615934371948242:A-9.830184341959654],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x20x96[-5.509540557861328,7.693107604980469:A-0.12716069182855183]],#1[T1s1x1x20x96[-0.6787744760513306,0.7704185843467712:A0.008161253841952506]]]))
<- ((),dict(cache_position:T7s1[20,20:A20.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x20x96[-5.509540557861328,7.693107604980469:A-0.12716069182855183]],#1[T1s1x1x20x96[-0.6787744760513306,0.7704185843467712:A0.008161253841952506]]]),input_ids:T7s1x1[25963,25963:A25963.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-17.643550872802734,6.164255142211914:A-8.786385558068053],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x21x96[-5.509540557861328,7.693107604980469:A-0.11984372260167941]],#1[T1s1x1x21x96[-0.6787744760513306,0.7704185843467712:A0.007688384226805774]]]))
<- ((),dict(cache_position:T7s1[21,21:A21.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x21x96[-5.509540557861328,7.693107604980469:A-0.11984372260167941]],#1[T1s1x1x21x96[-0.6787744760513306,0.7704185843467712:A0.007688384226805774]]]),input_ids:T7s1x1[29889,29889:A29889.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-14.255645751953125,7.926939010620117:A-7.896293622706086],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x22x96[-5.987019062042236,7.693107604980469:A-0.11909574120203836]],#1[T1s1x1x22x96[-0.6787744760513306,0.7704185843467712:A0.007864767629026766]]]))
<- ((),dict(cache_position:T7s1[22,22:A22.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x22x96[-5.987019062042236,7.693107604980469:A-0.11909574120203836]],#1[T1s1x1x22x96[-0.6787744760513306,0.7704185843467712:A0.007864767629026766]]]),input_ids:T7s1x1[13,13:A13.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-8.008821487426758,10.541921615600586:A-2.745266179916449],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x23x96[-5.987019062042236,7.693107604980469:A-0.1203865866913132]],#1[T1s1x1x23x96[-0.6787744760513306,0.7704185843467712:A0.008751925312520578]]]))
<- ((),dict(cache_position:T7s1[23,23:A23.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x23x96[-5.987019062042236,7.693107604980469:A-0.1203865866913132]],#1[T1s1x1x23x96[-0.6787744760513306,0.7704185843467712:A0.008751925312520578]]]),input_ids:T7s1x1[3492,3492:A3492.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-16.191024780273438,9.499390602111816:A-5.864637234960217],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x24x96[-5.987019062042236,7.693107604980469:A-0.12080627055037338]],#1[T1s1x1x24x96[-0.6787744760513306,0.7704185843467712:A0.008935451721343825]]]))
<- ((),dict(cache_position:T7s1[24,24:A24.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x24x96[-5.987019062042236,7.693107604980469:A-0.12080627055037338]],#1[T1s1x1x24x96[-0.6787744760513306,0.7704185843467712:A0.008935451721343825]]]),input_ids:T7s1x1[674,674:A674.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-16.663740158081055,9.625528335571289:A-6.569129421486053],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x25x96[-5.987019062042236,7.693107604980469:A-0.11726646279891914]],#1[T1s1x1x25x96[-0.6787744760513306,0.7704185843467712:A0.009409271212007297]]]))
<- ((),dict(cache_position:T7s1[25,25:A25.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x25x96[-5.987019062042236,7.693107604980469:A-0.11726646279891914]],#1[T1s1x1x25x96[-0.6787744760513306,0.7704185843467712:A0.009409271212007297]]]),input_ids:T7s1x1[748,748:A748.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-19.916133880615234,6.930554389953613:A-9.758149972834625],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x26x96[-5.987019062042236,7.693107604980469:A-0.11556891045951581]],#1[T1s1x1x26x96[-0.6787744760513306,0.7704185843467712:A0.009660339510114331]]]))
<- ((),dict(cache_position:T7s1[26,26:A26.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x26x96[-5.987019062042236,7.693107604980469:A-0.11556891045951581]],#1[T1s1x1x26x96[-0.6787744760513306,0.7704185843467712:A0.009660339510114331]]]),input_ids:T7s1x1[964,964:A964.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-20.35089111328125,6.057260036468506:A-9.874954451017082],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x27x96[-5.987019062042236,7.693107604980469:A-0.11108787408195411]],#1[T1s1x1x27x96[-0.6787744760513306,0.7704185843467712:A0.009390887548811911]]]))
<- ((),dict(cache_position:T7s1[27,27:A27.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x27x96[-5.987019062042236,7.693107604980469:A-0.11108787408195411]],#1[T1s1x1x27x96[-0.6787744760513306,0.7704185843467712:A0.009390887548811911]]]),input_ids:T7s1x1[697,697:A697.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-17.14153289794922,7.043888568878174:A-8.605550607407],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x28x96[-5.987019062042236,7.693107604980469:A-0.11001113096855468]],#1[T1s1x1x28x96[-0.6787744760513306,0.7704185843467712:A0.009548817436547367]]]))
<- ((),dict(cache_position:T7s1[28,28:A28.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x28x96[-5.987019062042236,7.693107604980469:A-0.11001113096855468]],#1[T1s1x1x28x96[-0.6787744760513306,0.7704185843467712:A0.009548817436547367]]]),input_ids:T7s1x1[297,297:A297.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-20.8441104888916,6.884699821472168:A-9.159340587275569],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x29x96[-5.987019062042236,7.70276403427124:A-0.10577449580899877]],#1[T1s1x1x29x96[-0.6787744760513306,0.7704185843467712:A0.009528502082905588]]]))
<- ((),dict(cache_position:T7s1[29,29:A29.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x29x96[-5.987019062042236,7.70276403427124:A-0.10577449580899877]],#1[T1s1x1x29x96[-0.6787744760513306,0.7704185843467712:A0.009528502082905588]]]),input_ids:T7s1x1[278,278:A278.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-19.315391540527344,4.614831924438477:A-9.002723350008484],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x30x96[-5.987019062042236,7.70276403427124:A-0.1008578367267344]],#1[T1s1x1x30x96[-0.6787744760513306,0.7704185843467712:A0.00966630508349024]]]))
<- ((),dict(cache_position:T7s1[30,30:A30.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x30x96[-5.987019062042236,7.70276403427124:A-0.1008578367267344]],#1[T1s1x1x30x96[-0.6787744760513306,0.7704185843467712:A0.00966630508349024]]]),input_ids:T7s1x1[29871,29871:A29871.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-11.507885932922363,15.988384246826172:A-2.5741930042644965],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x31x96[-5.987019062042236,7.70276403427124:A-0.09899786915278821]],#1[T1s1x1x31x96[-0.6787744760513306,0.7704185843467712:A0.008349076121806232]]]))
<- ((),dict(cache_position:T7s1[31,31:A31.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x31x96[-5.987019062042236,7.70276403427124:A-0.09899786915278821]],#1[T1s1x1x31x96[-0.6787744760513306,0.7704185843467712:A0.008349076121806232]]]),input_ids:T7s1x1[29896,29896:A29896.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-19.334497451782227,8.830238342285156:A-9.694990545970388],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x32x96[-5.987019062042236,7.70276403427124:A-0.09737615970952145]],#1[T1s1x1x32x96[-0.6787744760513306,0.7704185843467712:A0.008011651119844032]]]))
<- ((),dict(cache_position:T7s1[32,32:A32.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x32x96[-5.987019062042236,7.70276403427124:A-0.09737615970952145]],#1[T1s1x1x32x96[-0.6787744760513306,0.7704185843467712:A0.008011651119844032]]]),input_ids:T7s1x1[29946,29946:A29946.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-19.741653442382812,7.135903358459473:A-9.79210955552943],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x33x96[-5.987019062042236,7.70276403427124:A-0.09667280920841607]],#1[T1s1x1x33x96[-0.6787744760513306,0.7704185843467712:A0.007499413813465336]]]))
<- ((),dict(cache_position:T7s1[33,33:A33.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x33x96[-5.987019062042236,7.70276403427124:A-0.09667280920841607]],#1[T1s1x1x33x96[-0.6787744760513306,0.7704185843467712:A0.007499413813465336]]]),input_ids:T7s1x1[386,386:A386.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-22.21338653564453,4.775697231292725:A-10.594457698166371],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x34x96[-5.987019062042236,7.70276403427124:A-0.09190061861002452]],#1[T1s1x1x34x96[-0.6787744760513306,0.7704185843467712:A0.007487988068029225]]]))
<- ((),dict(cache_position:T7s1[34,34:A34.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x34x96[-5.987019062042236,7.70276403427124:A-0.09190061861002452]],#1[T1s1x1x34x96[-0.6787744760513306,0.7704185843467712:A0.007487988068029225]]]),input_ids:T7s1x1[322,322:A322.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-16.256813049316406,7.791941165924072:A-7.819804543957114],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x35x96[-5.987019062042236,7.70276403427124:A-0.09161416696439292]],#1[T1s1x1x35x96[-0.6787744760513306,0.7704185843467712:A0.00699816262567332]]]))
<- ((),dict(cache_position:T7s1[35,35:A35.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x35x96[-5.987019062042236,7.70276403427124:A-0.09161416696439292]],#1[T1s1x1x35x96[-0.6787744760513306,0.7704185843467712:A0.00699816262567332]]]),input_ids:T7s1x1[769,769:A769.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-15.366680145263672,6.7819390296936035:A-8.183435506382956],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x36x96[-5.987019062042236,7.70276403427124:A-0.09020163751157034]],#1[T1s1x1x36x96[-0.6787744760513306,0.7704185843467712:A0.0076035632615685245]]]))
<- ((),dict(cache_position:T7s1[36,36:A36.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x36x96[-5.987019062042236,7.70276403427124:A-0.09020163751157034]],#1[T1s1x1x36x96[-0.6787744760513306,0.7704185843467712:A0.0076035632615685245]]]),input_ids:T7s1x1[366,366:A366.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-16.212535858154297,10.488546371459961:A-6.096679806140484],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x37x96[-6.020578861236572,7.70276403427124:A-0.08767639670970295]],#1[T1s1x1x37x96[-0.6787744760513306,0.7704185843467712:A0.008052934940682922]]]))
<- ((),dict(cache_position:T7s1[37,37:A37.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x37x96[-6.020578861236572,7.70276403427124:A-0.08767639670970295]],#1[T1s1x1x37x96[-0.6787744760513306,0.7704185843467712:A0.008052934940682922]]]),input_ids:T7s1x1[925,925:A925.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-14.716682434082031,10.297311782836914:A-5.215678396251984],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x38x96[-6.020578861236572,7.70276403427124:A-0.08759201532765266]],#1[T1s1x1x38x96[-0.6787744760513306,0.7704185843467712:A0.007851843246255572]]]))
<- ((),dict(cache_position:T7s1[38,38:A38.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x38x96[-6.020578861236572,7.70276403427124:A-0.08759201532765266]],#1[T1s1x1x38x96[-0.6787744760513306,0.7704185843467712:A0.007851843246255572]]]),input_ids:T7s1x1[2125,2125:A2125.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-19.35562515258789,6.750692367553711:A-10.36610761905415],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x39x96[-6.020578861236572,7.70276403427124:A-0.08397989235922239]],#1[T1s1x1x39x96[-0.6787744760513306,0.7704185843467712:A0.007908436715961916]]]))
<- ((),dict(cache_position:T7s1[39,39:A39.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x39x96[-6.020578861236572,7.70276403427124:A-0.08397989235922239]],#1[T1s1x1x39x96[-0.6787744760513306,0.7704185843467712:A0.007908436715961916]]]),input_ids:T7s1x1[263,263:A263.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-18.591672897338867,7.181346893310547:A-7.389501655889908],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x40x96[-6.020578861236572,7.70276403427124:A-0.08059514915206213]],#1[T1s1x1x40x96[-0.6787744760513306,0.7704185843467712:A0.008126784816912884]]]))
<- ((),dict(cache_position:T7s1[40,40:A40.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x40x96[-6.020578861236572,7.70276403427124:A-0.08059514915206213]],#1[T1s1x1x40x96[-0.6787744760513306,0.7704185843467712:A0.008126784816912884]]]),input_ids:T7s1x1[29871,29871:A29871.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-7.655264377593994,18.418691635131836:A-0.8347461489182897],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x41x96[-6.020578861236572,7.70276403427124:A-0.07553431066684779]],#1[T1s1x1x41x96[-0.6787744760513306,0.7704185843467712:A0.00716838048653174]]]))
<- ((),dict(cache_position:T7s1[41,41:A41.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x41x96[-6.020578861236572,7.70276403427124:A-0.07553431066684779]],#1[T1s1x1x41x96[-0.6787744760513306,0.7704185843467712:A0.00716838048653174]]]),input_ids:T7s1x1[29945,29945:A29945.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-17.618488311767578,7.5059404373168945:A-9.185259233211168],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x42x96[-6.949457168579102,7.70276403427124:A-0.07523944943811027]],#1[T1s1x1x42x96[-0.6787744760513306,0.7704185843467712:A0.006704898665410448]]]))
<- ((),dict(cache_position:T7s1[42,42:A42.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x42x96[-6.949457168579102,7.70276403427124:A-0.07523944943811027]],#1[T1s1x1x42x96[-0.6787744760513306,0.7704185843467712:A0.006704898665410448]]]),input_ids:T7s1x1[29899,29899:A29899.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-12.02718448638916,11.478802680969238:A-4.218412199046463],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x43x96[-6.949457168579102,7.70276403427124:A-0.07125541826158148]],#1[T1s1x1x43x96[-0.6787744760513306,0.7704185843467712:A0.006634421506007566]]]))
<- ((),dict(cache_position:T7s1[43,43:A43.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x43x96[-6.949457168579102,7.70276403427124:A-0.07125541826158148]],#1[T1s1x1x43x96[-0.6787744760513306,0.7704185843467712:A0.006634421506007566]]]),input_ids:T7s1x1[29947,29947:A29947.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-19.648494720458984,4.7706780433654785:A-9.644542609302793],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x44x96[-6.949457168579102,7.70276403427124:A-0.0694073898192514]],#1[T1s1x1x44x96[-0.6787744760513306,0.7704185843467712:A0.0060485980257515785]]]))
<- ((),dict(cache_position:T7s1[44,44:A44.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x44x96[-6.949457168579102,7.70276403427124:A-0.0694073898192514]],#1[T1s1x1x44x96[-0.6787744760513306,0.7704185843467712:A0.0060485980257515785]]]),input_ids:T7s1x1[3275,3275:A3275.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-16.49040985107422,8.647275924682617:A-8.34354119547177],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x45x96[-6.949457168579102,7.70276403427124:A-0.06737759871636323]],#1[T1s1x1x45x96[-0.6787744760513306,0.7704185843467712:A0.005413111220989069]]]))
<- ((),dict(cache_position:T7s1[45,45:A45.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x45x96[-6.949457168579102,7.70276403427124:A-0.06737759871636323]],#1[T1s1x1x45x96[-0.6787744760513306,0.7704185843467712:A0.005413111220989069]]]),input_ids:T7s1x1[373,373:A373.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-17.80068016052246,6.506999969482422:A-10.12396313319495],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x46x96[-6.949457168579102,7.70276403427124:A-0.06422996340960958]],#1[T1s1x1x46x96[-0.6787744760513306,0.7704185843467712:A0.004871142359355002]]]))
<- ((),dict(cache_position:T7s1[46,46:A46.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x46x96[-6.949457168579102,7.70276403427124:A-0.06422996340960958]],#1[T1s1x1x46x96[-0.6787744760513306,0.7704185843467712:A0.004871142359355002]]]),input_ids:T7s1x1[278,278:A278.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-18.61307144165039,5.520223617553711:A-9.258860857699998],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x47x96[-6.949457168579102,7.70276403427124:A-0.06245628588304814]],#1[T1s1x1x47x96[-0.6787744760513306,0.7704185843467712:A0.005058194481505857]]]))
<- ((),dict(cache_position:T7s1[47,47:A47.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x47x96[-6.949457168579102,7.70276403427124:A-0.06245628588304814]],#1[T1s1x1x47x96[-0.6787744760513306,0.7704185843467712:A0.005058194481505857]]]),input_ids:T7s1x1[1473,1473:A1473.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-20.592666625976562,5.243607044219971:A-10.369299335102317],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x48x96[-6.949457168579102,7.70276403427124:A-0.06039970430060748]],#1[T1s1x1x48x96[-0.6787744760513306,0.7704185843467712:A0.004975992970151487]]]))
<- ((),dict(cache_position:T7s1[48,48:A48.0],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x48x96[-6.949457168579102,7.70276403427124:A-0.06039970430060748]],#1[T1s1x1x48x96[-0.6787744760513306,0.7704185843467712:A0.004975992970151487]]]),input_ids:T7s1x1[2814,2814:A2814.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> dict(logits:T1s1x1x32000[-12.931621551513672,11.199272155761719:A-5.162818993296707],past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x49x96[-6.949457168579102,7.70276403427124:A-0.05724061712005225]],#1[T1s1x1x49x96[-0.6787744760513306,0.7704185843467712:A0.004617188286386959]]]))
-- prompt Continue: it rains...
-- answer Continue: it rains...
It was the last of the first three times in the sixth.
You will go into one in the 14th and then you just take a 5-8 lead on the second leg,
Let’s restore the forward as it was.
model.forward = keep_model_forward
Another syntax with onnx_diagnostic.helpers.torch_test_helper.steel_forward().
with steel_forward(model):
model.generate(inputs, max_length=50, temperature=1, top_k=50, top_p=0.95, do_sample=True)
---- stolen forward for class LlamaForCausalLM -- iteration 0
<- args=() --- kwargs=dict(cache_position:T7s8,past_key_values:DynamicCache[serialized](#2[#0[],#0[]]),input_ids:T7s1x8,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x8x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x8x96],#1[T1s1x1x8x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 1
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x8x96],#1[T1s1x1x8x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x9x96],#1[T1s1x1x9x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 2
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x9x96],#1[T1s1x1x9x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x10x96],#1[T1s1x1x10x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 3
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x10x96],#1[T1s1x1x10x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x11x96],#1[T1s1x1x11x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 4
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x11x96],#1[T1s1x1x11x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x12x96],#1[T1s1x1x12x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 5
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x12x96],#1[T1s1x1x12x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x13x96],#1[T1s1x1x13x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 6
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x13x96],#1[T1s1x1x13x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x14x96],#1[T1s1x1x14x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 7
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x14x96],#1[T1s1x1x14x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x15x96],#1[T1s1x1x15x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 8
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x15x96],#1[T1s1x1x15x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x16x96],#1[T1s1x1x16x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 9
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x16x96],#1[T1s1x1x16x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x17x96],#1[T1s1x1x17x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 10
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x17x96],#1[T1s1x1x17x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x18x96],#1[T1s1x1x18x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 11
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x18x96],#1[T1s1x1x18x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x19x96],#1[T1s1x1x19x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 12
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x19x96],#1[T1s1x1x19x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x20x96],#1[T1s1x1x20x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 13
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x20x96],#1[T1s1x1x20x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x21x96],#1[T1s1x1x21x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 14
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x21x96],#1[T1s1x1x21x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x22x96],#1[T1s1x1x22x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 15
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x22x96],#1[T1s1x1x22x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x23x96],#1[T1s1x1x23x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 16
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x23x96],#1[T1s1x1x23x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x24x96],#1[T1s1x1x24x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 17
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x24x96],#1[T1s1x1x24x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x25x96],#1[T1s1x1x25x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 18
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x25x96],#1[T1s1x1x25x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x26x96],#1[T1s1x1x26x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 19
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x26x96],#1[T1s1x1x26x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x27x96],#1[T1s1x1x27x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 20
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x27x96],#1[T1s1x1x27x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x28x96],#1[T1s1x1x28x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 21
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x28x96],#1[T1s1x1x28x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x29x96],#1[T1s1x1x29x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 22
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x29x96],#1[T1s1x1x29x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x30x96],#1[T1s1x1x30x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 23
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x30x96],#1[T1s1x1x30x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x31x96],#1[T1s1x1x31x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 24
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x31x96],#1[T1s1x1x31x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x32x96],#1[T1s1x1x32x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 25
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x32x96],#1[T1s1x1x32x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x33x96],#1[T1s1x1x33x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 26
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x33x96],#1[T1s1x1x33x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x34x96],#1[T1s1x1x34x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 27
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x34x96],#1[T1s1x1x34x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x35x96],#1[T1s1x1x35x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 28
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x35x96],#1[T1s1x1x35x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x36x96],#1[T1s1x1x36x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 29
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x36x96],#1[T1s1x1x36x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x37x96],#1[T1s1x1x37x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 30
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x37x96],#1[T1s1x1x37x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x38x96],#1[T1s1x1x38x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 31
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x38x96],#1[T1s1x1x38x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x39x96],#1[T1s1x1x39x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 32
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x39x96],#1[T1s1x1x39x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x40x96],#1[T1s1x1x40x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 33
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x40x96],#1[T1s1x1x40x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x41x96],#1[T1s1x1x41x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 34
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x41x96],#1[T1s1x1x41x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x42x96],#1[T1s1x1x42x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 35
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x42x96],#1[T1s1x1x42x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x43x96],#1[T1s1x1x43x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 36
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x43x96],#1[T1s1x1x43x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x44x96],#1[T1s1x1x44x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 37
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x44x96],#1[T1s1x1x44x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x45x96],#1[T1s1x1x45x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 38
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x45x96],#1[T1s1x1x45x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x46x96],#1[T1s1x1x46x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 39
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x46x96],#1[T1s1x1x46x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x47x96],#1[T1s1x1x47x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 40
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x47x96],#1[T1s1x1x47x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x48x96],#1[T1s1x1x48x96]]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 41
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x48x96],#1[T1s1x1x48x96]]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> dict(logits:T1s1x1x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s1x1x49x96],#1[T1s1x1x49x96]]))
.
Untrained model¶
This part can skipped if you are only interested in exporting the original model. It is useful to create a unit test to ensure a specific architecture can be exported despite the many changes brought to torch or transformers.
Let’s create an untrained model using the config file provided
config.json
to create an untrained model:
onnx_diagnostic.torch_models.llms.get_tiny_llm().
Then let’s use it.
experiment = get_tiny_llm()
untrained_model, inputs, dynamic_shapes = (
experiment["model"],
experiment["inputs"],
experiment["dynamic_shapes"],
)
Before we run it, we make a copy of the inputs as the cache get modified by the execution. Then it is no longer valid associated with the previous input_ids and mask.
print("input type before", string_type(inputs, with_shape=True))
expected_output = untrained_model(**inputs)
print("input type after-", string_type(inputs, with_shape=True))
input type before dict(input_ids:T7s2x3,attention_mask:T7s2x33,position_ids:T7s2x3,past_key_values:DynamicCache[serialized](#2[#1[T1s2x1x30x96],#1[T1s2x1x30x96]]))
input type after- dict(input_ids:T7s2x3,attention_mask:T7s2x33,position_ids:T7s2x3,past_key_values:DynamicCache[serialized](#2[#1[T1s2x1x33x96],#1[T1s2x1x33x96]]))
The outputs
print("result type", string_type(expected_output, with_shape=True))
result type dict(logits:T1s2x3x32000,past_key_values:DynamicCache[serialized](#2[#1[T1s2x1x33x96],#1[T1s2x1x33x96]]))
It works.
ExportedProgram¶
try:
ep = torch.export.export(
untrained_model, (), kwargs=cloned_inputs, dynamic_shapes=dynamic_shapes, strict=False
)
print("It worked:")
print(ep)
except Exception as e:
# To work, it needs at least PRs:
# * https://github.com/huggingface/transformers/pull/36311
# * https://github.com/huggingface/transformers/pull/36652
print("It failed:", e)
It worked:
ExportedProgram:
class GraphModule(torch.nn.Module):
def forward(self, p_model_embed_tokens_weight: "f32[32000, 192]", p_model_layers_0_self_attn_q_proj_weight: "f32[192, 192]", p_model_layers_0_self_attn_k_proj_weight: "f32[96, 192]", p_model_layers_0_self_attn_v_proj_weight: "f32[96, 192]", p_model_layers_0_self_attn_o_proj_weight: "f32[192, 192]", p_model_layers_0_mlp_gate_proj_weight: "f32[1024, 192]", p_model_layers_0_mlp_up_proj_weight: "f32[1024, 192]", p_model_layers_0_mlp_down_proj_weight: "f32[192, 1024]", p_model_layers_0_input_layernorm_weight: "f32[192]", p_model_layers_0_post_attention_layernorm_weight: "f32[192]", p_model_norm_weight: "f32[192]", p_lm_head_weight: "f32[32000, 192]", b_model_rotary_emb_inv_freq: "f32[48]", input_ids: "i64[s41, s2]", attention_mask: "i64[s41, s2 + s67]", position_ids: "i64[s41, s2]", past_key_values_key_cache_0: "f32[s41, 1, s67, 96]", past_key_values_value_cache_0: "f32[s41, 1, s67, 96]"):
#
sym_size_int_22: "Sym(s41)" = torch.ops.aten.sym_size.int(input_ids, 0)
sym_size_int_23: "Sym(s2)" = torch.ops.aten.sym_size.int(input_ids, 1)
sym_size_int_24: "Sym(s67)" = torch.ops.aten.sym_size.int(past_key_values_key_cache_0, 2)
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:190 in forward, code: return F.embedding(
embedding: "f32[s41, s2, 192]" = torch.ops.aten.embedding.default(p_model_embed_tokens_weight, input_ids); p_model_embed_tokens_weight = input_ids = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:564 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
add: "Sym(s2 + s67)" = sym_size_int_24 + sym_size_int_23
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:563 in forward, code: cache_position = torch.arange(
arange: "i64[s2]" = torch.ops.aten.arange.start(sym_size_int_24, add, device = device(type='cpu'), pin_memory = False); sym_size_int_24 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:570 in forward, code: causal_mask = self._update_causal_mask(
full: "f32[s2, s2 + s67]" = torch.ops.aten.full.default([sym_size_int_23, add], -3.4028234663852886e+38, dtype = torch.float32, device = device(type='cpu'), pin_memory = False)
triu: "f32[s2, s2 + s67]" = torch.ops.aten.triu.default(full, 1); full = None
arange_1: "i64[s2 + s67]" = torch.ops.aten.arange.default(add, device = device(type='cpu'), pin_memory = False)
reshape: "i64[s2, 1]" = torch.ops.aten.reshape.default(arange, [-1, 1]); arange = None
gt: "b8[s2, s2 + s67]" = torch.ops.aten.gt.Tensor(arange_1, reshape); arange_1 = reshape = None
mul_: "f32[s2, s2 + s67]" = torch.ops.aten.mul_.Tensor(triu, gt); triu = gt = None
unsqueeze: "f32[1, s2, s2 + s67]" = torch.ops.aten.unsqueeze.default(mul_, 0); mul_ = None
unsqueeze_1: "f32[1, 1, s2, s2 + s67]" = torch.ops.aten.unsqueeze.default(unsqueeze, 1); unsqueeze = None
slice_1: "f32[1, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(unsqueeze_1, 2, 0, 9223372036854775807); unsqueeze_1 = None
slice_2: "f32[1, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_1, 3, 0, 9223372036854775807); slice_1 = None
expand: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.expand.default(slice_2, [sym_size_int_22, 1, -1, -1]); slice_2 = None
clone: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.clone.default(expand); expand = None
slice_3: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(clone)
slice_4: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_3, 1); slice_3 = None
slice_5: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_4, 2); slice_4 = None
slice_6: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_5, 3, None, add); slice_5 = None
slice_7: "i64[s41, s2 + s67]" = torch.ops.aten.slice.Tensor(attention_mask, 0, 0, 9223372036854775807); attention_mask = None
unsqueeze_2: "i64[s41, 1, s2 + s67]" = torch.ops.aten.unsqueeze.default(slice_7, 1); slice_7 = None
unsqueeze_3: "i64[s41, 1, 1, s2 + s67]" = torch.ops.aten.unsqueeze.default(unsqueeze_2, 2); unsqueeze_2 = None
slice_8: "i64[s41, 1, 1, s2 + s67]" = torch.ops.aten.slice.Tensor(unsqueeze_3, 3, 0, 9223372036854775807); unsqueeze_3 = None
_assert_tensor_metadata_default = torch.ops.aten._assert_tensor_metadata.default(slice_8, dtype = torch.int64, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default = None
to: "i64[s41, 1, 1, s2 + s67]" = torch.ops.aten.to.dtype_layout(slice_8, dtype = torch.int64, layout = torch.strided, device = device(type='cpu')); slice_8 = None
add_2: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.add.Tensor(slice_6, to); slice_6 = to = None
eq_4: "b8[s41, 1, s2, s2 + s67]" = torch.ops.aten.eq.Scalar(add_2, 0); add_2 = None
slice_9: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(clone)
slice_10: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_9, 1); slice_9 = None
slice_11: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_10, 2); slice_10 = None
slice_12: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_11, 3, None, add); slice_11 = None
masked_fill: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.masked_fill.Scalar(slice_12, eq_4, -3.4028234663852886e+38); slice_12 = eq_4 = None
slice_13: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(clone, 0, 0, 9223372036854775807)
slice_14: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_13, 1, 0, 9223372036854775807); slice_13 = None
slice_15: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_14, 2, 0, 9223372036854775807); slice_14 = None
copy_: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.copy_.default(slice_15, masked_fill); slice_15 = masked_fill = copy_ = None
# No stacktrace found for following nodes
submod_3 = self.submod_1
wrap_with_set_grad_enabled = torch.ops.higher_order.wrap_with_set_grad_enabled(False, submod_3, b_model_rotary_emb_inv_freq, sym_size_int_22, position_ids); submod_3 = b_model_rotary_emb_inv_freq = position_ids = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:152 in forward, code: return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
to_5: "f32[s41, s2, 96]" = wrap_with_set_grad_enabled[0]
to_6: "f32[s41, s2, 96]" = wrap_with_set_grad_enabled[1]; wrap_with_set_grad_enabled = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:80 in forward, code: hidden_states = hidden_states.to(torch.float32)
_assert_tensor_metadata_default_7 = torch.ops.aten._assert_tensor_metadata.default(embedding, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_7 = None
to_7: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(embedding, torch.float32); embedding = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:81 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_1: "f32[s41, s2, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_7, 2)
mean: "f32[s41, s2, 1]" = torch.ops.aten.mean.dim(pow_1, [-1], True); pow_1 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:82 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_3: "f32[s41, s2, 1]" = torch.ops.aten.add.Tensor(mean, 1e-05); mean = None
rsqrt: "f32[s41, s2, 1]" = torch.ops.aten.rsqrt.default(add_3); add_3 = None
mul_2: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(to_7, rsqrt); rsqrt = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:83 in forward, code: return self.weight * hidden_states.to(input_dtype)
_assert_tensor_metadata_default_8 = torch.ops.aten._assert_tensor_metadata.default(mul_2, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_8 = None
to_8: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(mul_2, torch.float32); mul_2 = None
mul_3: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(p_model_layers_0_input_layernorm_weight, to_8); p_model_layers_0_input_layernorm_weight = to_8 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear: "f32[s41, s2, 192]" = torch.ops.aten.linear.default(mul_3, p_model_layers_0_self_attn_q_proj_weight); p_model_layers_0_self_attn_q_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:281 in forward, code: query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view: "f32[s41, s2, 2, 96]" = torch.ops.aten.view.default(linear, [sym_size_int_22, sym_size_int_23, -1, 96]); linear = None
transpose_1: "f32[s41, 2, s2, 96]" = torch.ops.aten.transpose.int(view, 1, 2); view = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_1: "f32[s41, s2, 96]" = torch.ops.aten.linear.default(mul_3, p_model_layers_0_self_attn_k_proj_weight); p_model_layers_0_self_attn_k_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:282 in forward, code: key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view_1: "f32[s41, s2, 1, 96]" = torch.ops.aten.view.default(linear_1, [sym_size_int_22, sym_size_int_23, -1, 96]); linear_1 = None
transpose_2: "f32[s41, 1, s2, 96]" = torch.ops.aten.transpose.int(view_1, 1, 2); view_1 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_2: "f32[s41, s2, 96]" = torch.ops.aten.linear.default(mul_3, p_model_layers_0_self_attn_v_proj_weight); mul_3 = p_model_layers_0_self_attn_v_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:283 in forward, code: value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view_2: "f32[s41, s2, 1, 96]" = torch.ops.aten.view.default(linear_2, [sym_size_int_22, sym_size_int_23, -1, 96]); linear_2 = None
transpose_3: "f32[s41, 1, s2, 96]" = torch.ops.aten.transpose.int(view_2, 1, 2); view_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:286 in forward, code: query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
unsqueeze_7: "f32[s41, 1, s2, 96]" = torch.ops.aten.unsqueeze.default(to_5, 1); to_5 = None
unsqueeze_8: "f32[s41, 1, s2, 96]" = torch.ops.aten.unsqueeze.default(to_6, 1); to_6 = None
mul_4: "f32[s41, 2, s2, 96]" = torch.ops.aten.mul.Tensor(transpose_1, unsqueeze_7)
slice_19: "f32[s41, 2, s2, 48]" = torch.ops.aten.slice.Tensor(transpose_1, 3, 0, 48)
slice_20: "f32[s41, 2, s2, 48]" = torch.ops.aten.slice.Tensor(transpose_1, 3, 48, 9223372036854775807); transpose_1 = None
neg: "f32[s41, 2, s2, 48]" = torch.ops.aten.neg.default(slice_20); slice_20 = None
cat_1: "f32[s41, 2, s2, 96]" = torch.ops.aten.cat.default([neg, slice_19], -1); neg = slice_19 = None
mul_5: "f32[s41, 2, s2, 96]" = torch.ops.aten.mul.Tensor(cat_1, unsqueeze_8); cat_1 = None
add_4: "f32[s41, 2, s2, 96]" = torch.ops.aten.add.Tensor(mul_4, mul_5); mul_4 = mul_5 = None
mul_6: "f32[s41, 1, s2, 96]" = torch.ops.aten.mul.Tensor(transpose_2, unsqueeze_7); unsqueeze_7 = None
slice_21: "f32[s41, 1, s2, 48]" = torch.ops.aten.slice.Tensor(transpose_2, 3, 0, 48)
slice_22: "f32[s41, 1, s2, 48]" = torch.ops.aten.slice.Tensor(transpose_2, 3, 48, 9223372036854775807); transpose_2 = None
neg_1: "f32[s41, 1, s2, 48]" = torch.ops.aten.neg.default(slice_22); slice_22 = None
cat_2: "f32[s41, 1, s2, 96]" = torch.ops.aten.cat.default([neg_1, slice_21], -1); neg_1 = slice_21 = None
mul_7: "f32[s41, 1, s2, 96]" = torch.ops.aten.mul.Tensor(cat_2, unsqueeze_8); cat_2 = unsqueeze_8 = None
add_5: "f32[s41, 1, s2, 96]" = torch.ops.aten.add.Tensor(mul_6, mul_7); mul_6 = mul_7 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:291 in forward, code: key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
cat_3: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.cat.default([past_key_values_key_cache_0, add_5], -2); past_key_values_key_cache_0 = add_5 = None
cat_4: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.cat.default([past_key_values_value_cache_0, transpose_3], -2); past_key_values_value_cache_0 = transpose_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:303 in forward, code: attn_output, attn_weights = attention_interface(
slice_23: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(cat_3, 0, 0, 9223372036854775807)
slice_24: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(slice_23, 1, 0, 9223372036854775807); slice_23 = None
unsqueeze_9: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.unsqueeze.default(slice_24, 2); slice_24 = None
slice_25: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(unsqueeze_9, 3, 0, 9223372036854775807); unsqueeze_9 = None
slice_26: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(slice_25, 4, 0, 9223372036854775807); slice_25 = None
expand_2: "f32[s41, 1, 2, s2 + s67, 96]" = torch.ops.aten.expand.default(slice_26, [sym_size_int_22, 1, 2, add, 96]); slice_26 = None
reshape_1: "f32[s41, 2, s2 + s67, 96]" = torch.ops.aten.reshape.default(expand_2, [sym_size_int_22, 2, add, 96]); expand_2 = None
slice_27: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(cat_4, 0, 0, 9223372036854775807)
slice_28: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(slice_27, 1, 0, 9223372036854775807); slice_27 = None
unsqueeze_10: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.unsqueeze.default(slice_28, 2); slice_28 = None
slice_29: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(unsqueeze_10, 3, 0, 9223372036854775807); unsqueeze_10 = None
slice_30: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(slice_29, 4, 0, 9223372036854775807); slice_29 = None
expand_3: "f32[s41, 1, 2, s2 + s67, 96]" = torch.ops.aten.expand.default(slice_30, [sym_size_int_22, 1, 2, add, 96]); slice_30 = None
reshape_2: "f32[s41, 2, s2 + s67, 96]" = torch.ops.aten.reshape.default(expand_3, [sym_size_int_22, 2, add, 96]); expand_3 = None
slice_31: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(clone); clone = None
slice_32: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_31, 1); slice_31 = None
slice_33: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_32, 2); slice_32 = None
slice_34: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_33, 3, None, add); slice_33 = add = None
contiguous: "f32[s41, 2, s2, 96]" = torch.ops.aten.contiguous.default(add_4); add_4 = None
contiguous_1: "f32[s41, 2, s2 + s67, 96]" = torch.ops.aten.contiguous.default(reshape_1); reshape_1 = None
contiguous_2: "f32[s41, 2, s2 + s67, 96]" = torch.ops.aten.contiguous.default(reshape_2); reshape_2 = None
scaled_dot_product_attention: "f32[s41, 2, s2, 96]" = torch.ops.aten.scaled_dot_product_attention.default(contiguous, contiguous_1, contiguous_2, slice_34, scale = 0.10206207261596575); contiguous = contiguous_1 = contiguous_2 = slice_34 = None
transpose_4: "f32[s41, s2, 2, 96]" = torch.ops.aten.transpose.int(scaled_dot_product_attention, 1, 2); scaled_dot_product_attention = None
contiguous_3: "f32[s41, s2, 2, 96]" = torch.ops.aten.contiguous.default(transpose_4); transpose_4 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:314 in forward, code: attn_output = attn_output.reshape(*input_shape, -1).contiguous()
reshape_3: "f32[s41, s2, 192]" = torch.ops.aten.reshape.default(contiguous_3, [sym_size_int_22, sym_size_int_23, -1]); contiguous_3 = sym_size_int_22 = sym_size_int_23 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_3: "f32[s41, s2, 192]" = torch.ops.aten.linear.default(reshape_3, p_model_layers_0_self_attn_o_proj_weight); reshape_3 = p_model_layers_0_self_attn_o_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:358 in forward, code: hidden_states = residual + hidden_states
add_7: "f32[s41, s2, 192]" = torch.ops.aten.add.Tensor(to_7, linear_3); to_7 = linear_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:80 in forward, code: hidden_states = hidden_states.to(torch.float32)
_assert_tensor_metadata_default_9 = torch.ops.aten._assert_tensor_metadata.default(add_7, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_9 = None
to_9: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(add_7, torch.float32); add_7 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:81 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_2: "f32[s41, s2, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_9, 2)
mean_1: "f32[s41, s2, 1]" = torch.ops.aten.mean.dim(pow_2, [-1], True); pow_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:82 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_8: "f32[s41, s2, 1]" = torch.ops.aten.add.Tensor(mean_1, 1e-05); mean_1 = None
rsqrt_1: "f32[s41, s2, 1]" = torch.ops.aten.rsqrt.default(add_8); add_8 = None
mul_8: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(to_9, rsqrt_1); rsqrt_1 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:83 in forward, code: return self.weight * hidden_states.to(input_dtype)
_assert_tensor_metadata_default_10 = torch.ops.aten._assert_tensor_metadata.default(mul_8, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_10 = None
to_10: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(mul_8, torch.float32); mul_8 = None
mul_9: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(p_model_layers_0_post_attention_layernorm_weight, to_10); p_model_layers_0_post_attention_layernorm_weight = to_10 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_4: "f32[s41, s2, 1024]" = torch.ops.aten.linear.default(mul_9, p_model_layers_0_mlp_gate_proj_weight); p_model_layers_0_mlp_gate_proj_weight = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/activation.py:432 in forward, code: return F.silu(input, inplace=self.inplace)
silu: "f32[s41, s2, 1024]" = torch.ops.aten.silu.default(linear_4); linear_4 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_5: "f32[s41, s2, 1024]" = torch.ops.aten.linear.default(mul_9, p_model_layers_0_mlp_up_proj_weight); mul_9 = p_model_layers_0_mlp_up_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:201 in forward, code: down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
mul_10: "f32[s41, s2, 1024]" = torch.ops.aten.mul.Tensor(silu, linear_5); silu = linear_5 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_6: "f32[s41, s2, 192]" = torch.ops.aten.linear.default(mul_10, p_model_layers_0_mlp_down_proj_weight); mul_10 = p_model_layers_0_mlp_down_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:364 in forward, code: hidden_states = residual + hidden_states
add_9: "f32[s41, s2, 192]" = torch.ops.aten.add.Tensor(to_9, linear_6); to_9 = linear_6 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:80 in forward, code: hidden_states = hidden_states.to(torch.float32)
_assert_tensor_metadata_default_11 = torch.ops.aten._assert_tensor_metadata.default(add_9, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_11 = None
to_11: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(add_9, torch.float32); add_9 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:81 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_3: "f32[s41, s2, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_11, 2)
mean_2: "f32[s41, s2, 1]" = torch.ops.aten.mean.dim(pow_3, [-1], True); pow_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:82 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_10: "f32[s41, s2, 1]" = torch.ops.aten.add.Tensor(mean_2, 1e-05); mean_2 = None
rsqrt_2: "f32[s41, s2, 1]" = torch.ops.aten.rsqrt.default(add_10); add_10 = None
mul_11: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(to_11, rsqrt_2); to_11 = rsqrt_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:83 in forward, code: return self.weight * hidden_states.to(input_dtype)
_assert_tensor_metadata_default_12 = torch.ops.aten._assert_tensor_metadata.default(mul_11, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_12 = None
to_12: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(mul_11, torch.float32); mul_11 = None
mul_12: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(p_model_norm_weight, to_12); p_model_norm_weight = to_12 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:866 in forward, code: logits = self.lm_head(hidden_states[:, slice_indices, :])
slice_35: "f32[s41, s2, 192]" = torch.ops.aten.slice.Tensor(mul_12); mul_12 = None
slice_36: "f32[s41, s2, 192]" = torch.ops.aten.slice.Tensor(slice_35, 1, 0); slice_35 = None
slice_37: "f32[s41, s2, 192]" = torch.ops.aten.slice.Tensor(slice_36, 2); slice_36 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_7: "f32[s41, s2, 32000]" = torch.ops.aten.linear.default(slice_37, p_lm_head_weight); slice_37 = p_lm_head_weight = None
return (linear_7, cat_3, cat_4)
class submod_1(torch.nn.Module):
def forward(self, b_model_rotary_emb_inv_freq: "f32[48]", sym_size_int_22: "Sym(s41)", position_ids: "i64[s41, s2]"):
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:135 in forward, code: inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
unsqueeze_4: "f32[1, 48]" = torch.ops.aten.unsqueeze.default(b_model_rotary_emb_inv_freq, 0); b_model_rotary_emb_inv_freq = None
slice_16: "f32[1, 48]" = torch.ops.aten.slice.Tensor(unsqueeze_4, 1, 0, 9223372036854775807); unsqueeze_4 = None
unsqueeze_5: "f32[1, 48, 1]" = torch.ops.aten.unsqueeze.default(slice_16, 2); slice_16 = None
_assert_tensor_metadata_default_1 = torch.ops.aten._assert_tensor_metadata.default(unsqueeze_5, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_1 = None
to_1: "f32[1, 48, 1]" = torch.ops.aten.to.dtype(unsqueeze_5, torch.float32); unsqueeze_5 = None
expand_1: "f32[s41, 48, 1]" = torch.ops.aten.expand.default(to_1, [sym_size_int_22, -1, 1]); to_1 = sym_size_int_22 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:136 in forward, code: position_ids_expanded = position_ids[:, None, :].float()
slice_17: "i64[s41, s2]" = torch.ops.aten.slice.Tensor(position_ids, 0, 0, 9223372036854775807); position_ids = None
unsqueeze_6: "i64[s41, 1, s2]" = torch.ops.aten.unsqueeze.default(slice_17, 1); slice_17 = None
slice_18: "i64[s41, 1, s2]" = torch.ops.aten.slice.Tensor(unsqueeze_6, 2, 0, 9223372036854775807); unsqueeze_6 = None
_assert_tensor_metadata_default_2 = torch.ops.aten._assert_tensor_metadata.default(slice_18, dtype = torch.int64, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_2 = None
to_2: "f32[s41, 1, s2]" = torch.ops.aten.to.dtype(slice_18, torch.float32); slice_18 = None
# No stacktrace found for following nodes
submod_3 = self.submod_1
wrap_with_autocast = torch.ops.higher_order.wrap_with_autocast('cpu', torch.bfloat16, False, False, submod_3, expand_1, to_2); submod_3 = expand_1 = to_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:145 in forward, code: cos = emb.cos()
cos: "f32[s41, s2, 96]" = wrap_with_autocast[0]
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:146 in forward, code: sin = emb.sin()
sin: "f32[s41, s2, 96]" = wrap_with_autocast[1]; wrap_with_autocast = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:149 in forward, code: cos = cos * self.attention_scaling
mul: "f32[s41, s2, 96]" = torch.ops.aten.mul.Tensor(cos, 1.0); cos = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:150 in forward, code: sin = sin * self.attention_scaling
mul_1: "f32[s41, s2, 96]" = torch.ops.aten.mul.Tensor(sin, 1.0); sin = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:152 in forward, code: return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
_assert_tensor_metadata_default_5 = torch.ops.aten._assert_tensor_metadata.default(mul, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_5 = None
to_5: "f32[s41, s2, 96]" = torch.ops.aten.to.dtype(mul, torch.float32); mul = None
_assert_tensor_metadata_default_6 = torch.ops.aten._assert_tensor_metadata.default(mul_1, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_6 = None
to_6: "f32[s41, s2, 96]" = torch.ops.aten.to.dtype(mul_1, torch.float32); mul_1 = None
return (to_5, to_6)
class submod_1(torch.nn.Module):
def forward(self, expand_1: "f32[s41, 48, 1]", to_2: "f32[s41, 1, s2]"):
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:142 in forward, code: inv_freq_expanded.to(device=x.device, dtype=torch.float) @ position_ids_expanded.float()
_assert_tensor_metadata_default_3 = torch.ops.aten._assert_tensor_metadata.default(expand_1, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_3 = None
to_3: "f32[s41, 48, 1]" = torch.ops.aten.to.device(expand_1, device(type='cpu'), torch.float32); expand_1 = None
_assert_tensor_metadata_default_4 = torch.ops.aten._assert_tensor_metadata.default(to_2, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_4 = None
to_4: "f32[s41, 1, s2]" = torch.ops.aten.to.dtype(to_2, torch.float32); to_2 = None
matmul: "f32[s41, 48, s2]" = torch.ops.aten.matmul.default(to_3, to_4); to_3 = to_4 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:143 in forward, code: ).transpose(1, 2)
transpose: "f32[s41, s2, 48]" = torch.ops.aten.transpose.int(matmul, 1, 2); matmul = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:144 in forward, code: emb = torch.cat((freqs, freqs), dim=-1)
cat: "f32[s41, s2, 96]" = torch.ops.aten.cat.default([transpose, transpose], -1); transpose = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:145 in forward, code: cos = emb.cos()
cos: "f32[s41, s2, 96]" = torch.ops.aten.cos.default(cat)
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:146 in forward, code: sin = emb.sin()
sin: "f32[s41, s2, 96]" = torch.ops.aten.sin.default(cat); cat = None
return (cos, sin)
Graph signature:
# inputs
p_model_embed_tokens_weight: PARAMETER target='model.embed_tokens.weight'
p_model_layers_0_self_attn_q_proj_weight: PARAMETER target='model.layers.0.self_attn.q_proj.weight'
p_model_layers_0_self_attn_k_proj_weight: PARAMETER target='model.layers.0.self_attn.k_proj.weight'
p_model_layers_0_self_attn_v_proj_weight: PARAMETER target='model.layers.0.self_attn.v_proj.weight'
p_model_layers_0_self_attn_o_proj_weight: PARAMETER target='model.layers.0.self_attn.o_proj.weight'
p_model_layers_0_mlp_gate_proj_weight: PARAMETER target='model.layers.0.mlp.gate_proj.weight'
p_model_layers_0_mlp_up_proj_weight: PARAMETER target='model.layers.0.mlp.up_proj.weight'
p_model_layers_0_mlp_down_proj_weight: PARAMETER target='model.layers.0.mlp.down_proj.weight'
p_model_layers_0_input_layernorm_weight: PARAMETER target='model.layers.0.input_layernorm.weight'
p_model_layers_0_post_attention_layernorm_weight: PARAMETER target='model.layers.0.post_attention_layernorm.weight'
p_model_norm_weight: PARAMETER target='model.norm.weight'
p_lm_head_weight: PARAMETER target='lm_head.weight'
b_model_rotary_emb_inv_freq: BUFFER target='model.rotary_emb.inv_freq' persistent=False
input_ids: USER_INPUT
attention_mask: USER_INPUT
position_ids: USER_INPUT
past_key_values_key_cache_0: USER_INPUT
past_key_values_value_cache_0: USER_INPUT
# outputs
linear_7: USER_OUTPUT
cat_3: USER_OUTPUT
cat_4: USER_OUTPUT
Range constraints: {s41: VR[1, 1024], s2: VR[2, 4096], s2 + s67: VR[4, 8192], s67: VR[1, 4096]}
Back to the original model¶
Let’s use the same dummy inputs but we use the downloaded model.
Dummy inputs and dynamic shapes are created by function
onnx_diagnostic.torch_models.llms.get_tiny_llm().
data = get_tiny_llm()
inputs, dynamic_shapes = data["inputs"], data["dynamic_shapes"]
Let’s print the inputs.
print(string_type(inputs, with_shape=True))
dict(input_ids:T7s2x3,attention_mask:T7s2x33,position_ids:T7s2x3,past_key_values:DynamicCache[serialized](#2[#1[T1s2x1x30x96],#1[T1s2x1x30x96]]))
{'attention_mask': {0: Dim('batch', min=1, max=1024),
1: _DimHint(type=<_DimHintType.DYNAMIC: 3>,
min=None,
max=None,
_factory=True)},
'input_ids': {0: Dim('batch', min=1, max=1024),
1: Dim('seq_length', min=1, max=4096)},
'past_key_values': [[{0: Dim('batch', min=1, max=1024),
2: Dim('cache_length', min=1, max=4096)}],
[{0: Dim('batch', min=1, max=1024),
2: Dim('cache_length', min=1, max=4096)}]],
'position_ids': {0: Dim('batch', min=1, max=1024),
1: _DimHint(type=<_DimHintType.DYNAMIC: 3>,
min=None,
max=None,
_factory=True)}}
And Let’s finally export.
try:
ep = torch.export.export(
model, (), kwargs=cloned_inputs, dynamic_shapes=dynamic_shapes, strict=False
)
print("It worked:")
print(ep)
except Exception as e:
# To work, it needs at least PRs:
# * https://github.com/huggingface/transformers/pull/36311
# * https://github.com/huggingface/transformers/pull/36652
print("It failed:", e)
It worked:
ExportedProgram:
class GraphModule(torch.nn.Module):
def forward(self, p_model_embed_tokens_weight: "f32[32000, 192]", p_model_layers_0_self_attn_q_proj_weight: "f32[192, 192]", p_model_layers_0_self_attn_k_proj_weight: "f32[96, 192]", p_model_layers_0_self_attn_v_proj_weight: "f32[96, 192]", p_model_layers_0_self_attn_o_proj_weight: "f32[192, 192]", p_model_layers_0_mlp_gate_proj_weight: "f32[1024, 192]", p_model_layers_0_mlp_up_proj_weight: "f32[1024, 192]", p_model_layers_0_mlp_down_proj_weight: "f32[192, 1024]", p_model_layers_0_input_layernorm_weight: "f32[192]", p_model_layers_0_post_attention_layernorm_weight: "f32[192]", p_model_norm_weight: "f32[192]", p_lm_head_weight: "f32[32000, 192]", b_model_rotary_emb_inv_freq: "f32[48]", input_ids: "i64[s41, s2]", attention_mask: "i64[s41, s2 + s67]", position_ids: "i64[s41, s2]", past_key_values_key_cache_0: "f32[s41, 1, s67, 96]", past_key_values_value_cache_0: "f32[s41, 1, s67, 96]"):
#
sym_size_int_22: "Sym(s41)" = torch.ops.aten.sym_size.int(input_ids, 0)
sym_size_int_23: "Sym(s2)" = torch.ops.aten.sym_size.int(input_ids, 1)
sym_size_int_24: "Sym(s67)" = torch.ops.aten.sym_size.int(past_key_values_key_cache_0, 2)
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:190 in forward, code: return F.embedding(
embedding: "f32[s41, s2, 192]" = torch.ops.aten.embedding.default(p_model_embed_tokens_weight, input_ids); p_model_embed_tokens_weight = input_ids = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:564 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
add: "Sym(s2 + s67)" = sym_size_int_24 + sym_size_int_23
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:563 in forward, code: cache_position = torch.arange(
arange: "i64[s2]" = torch.ops.aten.arange.start(sym_size_int_24, add, device = device(type='cpu'), pin_memory = False); sym_size_int_24 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:570 in forward, code: causal_mask = self._update_causal_mask(
full: "f32[s2, s2 + s67]" = torch.ops.aten.full.default([sym_size_int_23, add], -3.4028234663852886e+38, dtype = torch.float32, device = device(type='cpu'), pin_memory = False)
triu: "f32[s2, s2 + s67]" = torch.ops.aten.triu.default(full, 1); full = None
arange_1: "i64[s2 + s67]" = torch.ops.aten.arange.default(add, device = device(type='cpu'), pin_memory = False)
reshape: "i64[s2, 1]" = torch.ops.aten.reshape.default(arange, [-1, 1]); arange = None
gt: "b8[s2, s2 + s67]" = torch.ops.aten.gt.Tensor(arange_1, reshape); arange_1 = reshape = None
mul_: "f32[s2, s2 + s67]" = torch.ops.aten.mul_.Tensor(triu, gt); triu = gt = None
unsqueeze: "f32[1, s2, s2 + s67]" = torch.ops.aten.unsqueeze.default(mul_, 0); mul_ = None
unsqueeze_1: "f32[1, 1, s2, s2 + s67]" = torch.ops.aten.unsqueeze.default(unsqueeze, 1); unsqueeze = None
slice_1: "f32[1, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(unsqueeze_1, 2, 0, 9223372036854775807); unsqueeze_1 = None
slice_2: "f32[1, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_1, 3, 0, 9223372036854775807); slice_1 = None
expand: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.expand.default(slice_2, [sym_size_int_22, 1, -1, -1]); slice_2 = None
clone: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.clone.default(expand); expand = None
slice_3: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(clone)
slice_4: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_3, 1); slice_3 = None
slice_5: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_4, 2); slice_4 = None
slice_6: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_5, 3, None, add); slice_5 = None
slice_7: "i64[s41, s2 + s67]" = torch.ops.aten.slice.Tensor(attention_mask, 0, 0, 9223372036854775807); attention_mask = None
unsqueeze_2: "i64[s41, 1, s2 + s67]" = torch.ops.aten.unsqueeze.default(slice_7, 1); slice_7 = None
unsqueeze_3: "i64[s41, 1, 1, s2 + s67]" = torch.ops.aten.unsqueeze.default(unsqueeze_2, 2); unsqueeze_2 = None
slice_8: "i64[s41, 1, 1, s2 + s67]" = torch.ops.aten.slice.Tensor(unsqueeze_3, 3, 0, 9223372036854775807); unsqueeze_3 = None
_assert_tensor_metadata_default = torch.ops.aten._assert_tensor_metadata.default(slice_8, dtype = torch.int64, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default = None
to: "i64[s41, 1, 1, s2 + s67]" = torch.ops.aten.to.dtype_layout(slice_8, dtype = torch.int64, layout = torch.strided, device = device(type='cpu')); slice_8 = None
add_2: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.add.Tensor(slice_6, to); slice_6 = to = None
eq_4: "b8[s41, 1, s2, s2 + s67]" = torch.ops.aten.eq.Scalar(add_2, 0); add_2 = None
slice_9: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(clone)
slice_10: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_9, 1); slice_9 = None
slice_11: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_10, 2); slice_10 = None
slice_12: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_11, 3, None, add); slice_11 = None
masked_fill: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.masked_fill.Scalar(slice_12, eq_4, -3.4028234663852886e+38); slice_12 = eq_4 = None
slice_13: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(clone, 0, 0, 9223372036854775807)
slice_14: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_13, 1, 0, 9223372036854775807); slice_13 = None
slice_15: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_14, 2, 0, 9223372036854775807); slice_14 = None
copy_: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.copy_.default(slice_15, masked_fill); slice_15 = masked_fill = copy_ = None
# No stacktrace found for following nodes
submod_3 = self.submod_1
wrap_with_set_grad_enabled = torch.ops.higher_order.wrap_with_set_grad_enabled(False, submod_3, b_model_rotary_emb_inv_freq, sym_size_int_22, position_ids); submod_3 = b_model_rotary_emb_inv_freq = position_ids = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:152 in forward, code: return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
to_5: "f32[s41, s2, 96]" = wrap_with_set_grad_enabled[0]
to_6: "f32[s41, s2, 96]" = wrap_with_set_grad_enabled[1]; wrap_with_set_grad_enabled = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:80 in forward, code: hidden_states = hidden_states.to(torch.float32)
_assert_tensor_metadata_default_7 = torch.ops.aten._assert_tensor_metadata.default(embedding, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_7 = None
to_7: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(embedding, torch.float32); embedding = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:81 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_1: "f32[s41, s2, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_7, 2)
mean: "f32[s41, s2, 1]" = torch.ops.aten.mean.dim(pow_1, [-1], True); pow_1 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:82 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_3: "f32[s41, s2, 1]" = torch.ops.aten.add.Tensor(mean, 1e-05); mean = None
rsqrt: "f32[s41, s2, 1]" = torch.ops.aten.rsqrt.default(add_3); add_3 = None
mul_2: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(to_7, rsqrt); rsqrt = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:83 in forward, code: return self.weight * hidden_states.to(input_dtype)
_assert_tensor_metadata_default_8 = torch.ops.aten._assert_tensor_metadata.default(mul_2, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_8 = None
to_8: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(mul_2, torch.float32); mul_2 = None
mul_3: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(p_model_layers_0_input_layernorm_weight, to_8); p_model_layers_0_input_layernorm_weight = to_8 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear: "f32[s41, s2, 192]" = torch.ops.aten.linear.default(mul_3, p_model_layers_0_self_attn_q_proj_weight); p_model_layers_0_self_attn_q_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:281 in forward, code: query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view: "f32[s41, s2, 2, 96]" = torch.ops.aten.view.default(linear, [sym_size_int_22, sym_size_int_23, -1, 96]); linear = None
transpose_1: "f32[s41, 2, s2, 96]" = torch.ops.aten.transpose.int(view, 1, 2); view = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_1: "f32[s41, s2, 96]" = torch.ops.aten.linear.default(mul_3, p_model_layers_0_self_attn_k_proj_weight); p_model_layers_0_self_attn_k_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:282 in forward, code: key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view_1: "f32[s41, s2, 1, 96]" = torch.ops.aten.view.default(linear_1, [sym_size_int_22, sym_size_int_23, -1, 96]); linear_1 = None
transpose_2: "f32[s41, 1, s2, 96]" = torch.ops.aten.transpose.int(view_1, 1, 2); view_1 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_2: "f32[s41, s2, 96]" = torch.ops.aten.linear.default(mul_3, p_model_layers_0_self_attn_v_proj_weight); mul_3 = p_model_layers_0_self_attn_v_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:283 in forward, code: value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view_2: "f32[s41, s2, 1, 96]" = torch.ops.aten.view.default(linear_2, [sym_size_int_22, sym_size_int_23, -1, 96]); linear_2 = None
transpose_3: "f32[s41, 1, s2, 96]" = torch.ops.aten.transpose.int(view_2, 1, 2); view_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:286 in forward, code: query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
unsqueeze_7: "f32[s41, 1, s2, 96]" = torch.ops.aten.unsqueeze.default(to_5, 1); to_5 = None
unsqueeze_8: "f32[s41, 1, s2, 96]" = torch.ops.aten.unsqueeze.default(to_6, 1); to_6 = None
mul_4: "f32[s41, 2, s2, 96]" = torch.ops.aten.mul.Tensor(transpose_1, unsqueeze_7)
slice_19: "f32[s41, 2, s2, 48]" = torch.ops.aten.slice.Tensor(transpose_1, 3, 0, 48)
slice_20: "f32[s41, 2, s2, 48]" = torch.ops.aten.slice.Tensor(transpose_1, 3, 48, 9223372036854775807); transpose_1 = None
neg: "f32[s41, 2, s2, 48]" = torch.ops.aten.neg.default(slice_20); slice_20 = None
cat_1: "f32[s41, 2, s2, 96]" = torch.ops.aten.cat.default([neg, slice_19], -1); neg = slice_19 = None
mul_5: "f32[s41, 2, s2, 96]" = torch.ops.aten.mul.Tensor(cat_1, unsqueeze_8); cat_1 = None
add_4: "f32[s41, 2, s2, 96]" = torch.ops.aten.add.Tensor(mul_4, mul_5); mul_4 = mul_5 = None
mul_6: "f32[s41, 1, s2, 96]" = torch.ops.aten.mul.Tensor(transpose_2, unsqueeze_7); unsqueeze_7 = None
slice_21: "f32[s41, 1, s2, 48]" = torch.ops.aten.slice.Tensor(transpose_2, 3, 0, 48)
slice_22: "f32[s41, 1, s2, 48]" = torch.ops.aten.slice.Tensor(transpose_2, 3, 48, 9223372036854775807); transpose_2 = None
neg_1: "f32[s41, 1, s2, 48]" = torch.ops.aten.neg.default(slice_22); slice_22 = None
cat_2: "f32[s41, 1, s2, 96]" = torch.ops.aten.cat.default([neg_1, slice_21], -1); neg_1 = slice_21 = None
mul_7: "f32[s41, 1, s2, 96]" = torch.ops.aten.mul.Tensor(cat_2, unsqueeze_8); cat_2 = unsqueeze_8 = None
add_5: "f32[s41, 1, s2, 96]" = torch.ops.aten.add.Tensor(mul_6, mul_7); mul_6 = mul_7 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:291 in forward, code: key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
cat_3: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.cat.default([past_key_values_key_cache_0, add_5], -2); past_key_values_key_cache_0 = add_5 = None
cat_4: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.cat.default([past_key_values_value_cache_0, transpose_3], -2); past_key_values_value_cache_0 = transpose_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:303 in forward, code: attn_output, attn_weights = attention_interface(
slice_23: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(cat_3, 0, 0, 9223372036854775807)
slice_24: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(slice_23, 1, 0, 9223372036854775807); slice_23 = None
unsqueeze_9: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.unsqueeze.default(slice_24, 2); slice_24 = None
slice_25: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(unsqueeze_9, 3, 0, 9223372036854775807); unsqueeze_9 = None
slice_26: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(slice_25, 4, 0, 9223372036854775807); slice_25 = None
expand_2: "f32[s41, 1, 2, s2 + s67, 96]" = torch.ops.aten.expand.default(slice_26, [sym_size_int_22, 1, 2, add, 96]); slice_26 = None
reshape_1: "f32[s41, 2, s2 + s67, 96]" = torch.ops.aten.reshape.default(expand_2, [sym_size_int_22, 2, add, 96]); expand_2 = None
slice_27: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(cat_4, 0, 0, 9223372036854775807)
slice_28: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(slice_27, 1, 0, 9223372036854775807); slice_27 = None
unsqueeze_10: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.unsqueeze.default(slice_28, 2); slice_28 = None
slice_29: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(unsqueeze_10, 3, 0, 9223372036854775807); unsqueeze_10 = None
slice_30: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(slice_29, 4, 0, 9223372036854775807); slice_29 = None
expand_3: "f32[s41, 1, 2, s2 + s67, 96]" = torch.ops.aten.expand.default(slice_30, [sym_size_int_22, 1, 2, add, 96]); slice_30 = None
reshape_2: "f32[s41, 2, s2 + s67, 96]" = torch.ops.aten.reshape.default(expand_3, [sym_size_int_22, 2, add, 96]); expand_3 = None
slice_31: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(clone); clone = None
slice_32: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_31, 1); slice_31 = None
slice_33: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_32, 2); slice_32 = None
slice_34: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_33, 3, None, add); slice_33 = add = None
contiguous: "f32[s41, 2, s2, 96]" = torch.ops.aten.contiguous.default(add_4); add_4 = None
contiguous_1: "f32[s41, 2, s2 + s67, 96]" = torch.ops.aten.contiguous.default(reshape_1); reshape_1 = None
contiguous_2: "f32[s41, 2, s2 + s67, 96]" = torch.ops.aten.contiguous.default(reshape_2); reshape_2 = None
scaled_dot_product_attention: "f32[s41, 2, s2, 96]" = torch.ops.aten.scaled_dot_product_attention.default(contiguous, contiguous_1, contiguous_2, slice_34, scale = 0.10206207261596575); contiguous = contiguous_1 = contiguous_2 = slice_34 = None
transpose_4: "f32[s41, s2, 2, 96]" = torch.ops.aten.transpose.int(scaled_dot_product_attention, 1, 2); scaled_dot_product_attention = None
contiguous_3: "f32[s41, s2, 2, 96]" = torch.ops.aten.contiguous.default(transpose_4); transpose_4 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:314 in forward, code: attn_output = attn_output.reshape(*input_shape, -1).contiguous()
reshape_3: "f32[s41, s2, 192]" = torch.ops.aten.reshape.default(contiguous_3, [sym_size_int_22, sym_size_int_23, -1]); contiguous_3 = sym_size_int_22 = sym_size_int_23 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_3: "f32[s41, s2, 192]" = torch.ops.aten.linear.default(reshape_3, p_model_layers_0_self_attn_o_proj_weight); reshape_3 = p_model_layers_0_self_attn_o_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:358 in forward, code: hidden_states = residual + hidden_states
add_7: "f32[s41, s2, 192]" = torch.ops.aten.add.Tensor(to_7, linear_3); to_7 = linear_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:80 in forward, code: hidden_states = hidden_states.to(torch.float32)
_assert_tensor_metadata_default_9 = torch.ops.aten._assert_tensor_metadata.default(add_7, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_9 = None
to_9: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(add_7, torch.float32); add_7 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:81 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_2: "f32[s41, s2, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_9, 2)
mean_1: "f32[s41, s2, 1]" = torch.ops.aten.mean.dim(pow_2, [-1], True); pow_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:82 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_8: "f32[s41, s2, 1]" = torch.ops.aten.add.Tensor(mean_1, 1e-05); mean_1 = None
rsqrt_1: "f32[s41, s2, 1]" = torch.ops.aten.rsqrt.default(add_8); add_8 = None
mul_8: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(to_9, rsqrt_1); rsqrt_1 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:83 in forward, code: return self.weight * hidden_states.to(input_dtype)
_assert_tensor_metadata_default_10 = torch.ops.aten._assert_tensor_metadata.default(mul_8, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_10 = None
to_10: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(mul_8, torch.float32); mul_8 = None
mul_9: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(p_model_layers_0_post_attention_layernorm_weight, to_10); p_model_layers_0_post_attention_layernorm_weight = to_10 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_4: "f32[s41, s2, 1024]" = torch.ops.aten.linear.default(mul_9, p_model_layers_0_mlp_gate_proj_weight); p_model_layers_0_mlp_gate_proj_weight = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/activation.py:432 in forward, code: return F.silu(input, inplace=self.inplace)
silu: "f32[s41, s2, 1024]" = torch.ops.aten.silu.default(linear_4); linear_4 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_5: "f32[s41, s2, 1024]" = torch.ops.aten.linear.default(mul_9, p_model_layers_0_mlp_up_proj_weight); mul_9 = p_model_layers_0_mlp_up_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:201 in forward, code: down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
mul_10: "f32[s41, s2, 1024]" = torch.ops.aten.mul.Tensor(silu, linear_5); silu = linear_5 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_6: "f32[s41, s2, 192]" = torch.ops.aten.linear.default(mul_10, p_model_layers_0_mlp_down_proj_weight); mul_10 = p_model_layers_0_mlp_down_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:364 in forward, code: hidden_states = residual + hidden_states
add_9: "f32[s41, s2, 192]" = torch.ops.aten.add.Tensor(to_9, linear_6); to_9 = linear_6 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:80 in forward, code: hidden_states = hidden_states.to(torch.float32)
_assert_tensor_metadata_default_11 = torch.ops.aten._assert_tensor_metadata.default(add_9, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_11 = None
to_11: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(add_9, torch.float32); add_9 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:81 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_3: "f32[s41, s2, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_11, 2)
mean_2: "f32[s41, s2, 1]" = torch.ops.aten.mean.dim(pow_3, [-1], True); pow_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:82 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_10: "f32[s41, s2, 1]" = torch.ops.aten.add.Tensor(mean_2, 1e-05); mean_2 = None
rsqrt_2: "f32[s41, s2, 1]" = torch.ops.aten.rsqrt.default(add_10); add_10 = None
mul_11: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(to_11, rsqrt_2); to_11 = rsqrt_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:83 in forward, code: return self.weight * hidden_states.to(input_dtype)
_assert_tensor_metadata_default_12 = torch.ops.aten._assert_tensor_metadata.default(mul_11, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_12 = None
to_12: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(mul_11, torch.float32); mul_11 = None
mul_12: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(p_model_norm_weight, to_12); p_model_norm_weight = to_12 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:866 in forward, code: logits = self.lm_head(hidden_states[:, slice_indices, :])
slice_35: "f32[s41, s2, 192]" = torch.ops.aten.slice.Tensor(mul_12); mul_12 = None
slice_36: "f32[s41, s2, 192]" = torch.ops.aten.slice.Tensor(slice_35, 1, 0); slice_35 = None
slice_37: "f32[s41, s2, 192]" = torch.ops.aten.slice.Tensor(slice_36, 2); slice_36 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_7: "f32[s41, s2, 32000]" = torch.ops.aten.linear.default(slice_37, p_lm_head_weight); slice_37 = p_lm_head_weight = None
return (linear_7, cat_3, cat_4)
class submod_1(torch.nn.Module):
def forward(self, b_model_rotary_emb_inv_freq: "f32[48]", sym_size_int_22: "Sym(s41)", position_ids: "i64[s41, s2]"):
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:135 in forward, code: inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
unsqueeze_4: "f32[1, 48]" = torch.ops.aten.unsqueeze.default(b_model_rotary_emb_inv_freq, 0); b_model_rotary_emb_inv_freq = None
slice_16: "f32[1, 48]" = torch.ops.aten.slice.Tensor(unsqueeze_4, 1, 0, 9223372036854775807); unsqueeze_4 = None
unsqueeze_5: "f32[1, 48, 1]" = torch.ops.aten.unsqueeze.default(slice_16, 2); slice_16 = None
_assert_tensor_metadata_default_1 = torch.ops.aten._assert_tensor_metadata.default(unsqueeze_5, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_1 = None
to_1: "f32[1, 48, 1]" = torch.ops.aten.to.dtype(unsqueeze_5, torch.float32); unsqueeze_5 = None
expand_1: "f32[s41, 48, 1]" = torch.ops.aten.expand.default(to_1, [sym_size_int_22, -1, 1]); to_1 = sym_size_int_22 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:136 in forward, code: position_ids_expanded = position_ids[:, None, :].float()
slice_17: "i64[s41, s2]" = torch.ops.aten.slice.Tensor(position_ids, 0, 0, 9223372036854775807); position_ids = None
unsqueeze_6: "i64[s41, 1, s2]" = torch.ops.aten.unsqueeze.default(slice_17, 1); slice_17 = None
slice_18: "i64[s41, 1, s2]" = torch.ops.aten.slice.Tensor(unsqueeze_6, 2, 0, 9223372036854775807); unsqueeze_6 = None
_assert_tensor_metadata_default_2 = torch.ops.aten._assert_tensor_metadata.default(slice_18, dtype = torch.int64, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_2 = None
to_2: "f32[s41, 1, s2]" = torch.ops.aten.to.dtype(slice_18, torch.float32); slice_18 = None
# No stacktrace found for following nodes
submod_3 = self.submod_1
wrap_with_autocast = torch.ops.higher_order.wrap_with_autocast('cpu', torch.bfloat16, False, False, submod_3, expand_1, to_2); submod_3 = expand_1 = to_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:145 in forward, code: cos = emb.cos()
cos: "f32[s41, s2, 96]" = wrap_with_autocast[0]
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:146 in forward, code: sin = emb.sin()
sin: "f32[s41, s2, 96]" = wrap_with_autocast[1]; wrap_with_autocast = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:149 in forward, code: cos = cos * self.attention_scaling
mul: "f32[s41, s2, 96]" = torch.ops.aten.mul.Tensor(cos, 1.0); cos = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:150 in forward, code: sin = sin * self.attention_scaling
mul_1: "f32[s41, s2, 96]" = torch.ops.aten.mul.Tensor(sin, 1.0); sin = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:152 in forward, code: return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
_assert_tensor_metadata_default_5 = torch.ops.aten._assert_tensor_metadata.default(mul, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_5 = None
to_5: "f32[s41, s2, 96]" = torch.ops.aten.to.dtype(mul, torch.float32); mul = None
_assert_tensor_metadata_default_6 = torch.ops.aten._assert_tensor_metadata.default(mul_1, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_6 = None
to_6: "f32[s41, s2, 96]" = torch.ops.aten.to.dtype(mul_1, torch.float32); mul_1 = None
return (to_5, to_6)
class submod_1(torch.nn.Module):
def forward(self, expand_1: "f32[s41, 48, 1]", to_2: "f32[s41, 1, s2]"):
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:142 in forward, code: inv_freq_expanded.to(device=x.device, dtype=torch.float) @ position_ids_expanded.float()
_assert_tensor_metadata_default_3 = torch.ops.aten._assert_tensor_metadata.default(expand_1, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_3 = None
to_3: "f32[s41, 48, 1]" = torch.ops.aten.to.device(expand_1, device(type='cpu'), torch.float32); expand_1 = None
_assert_tensor_metadata_default_4 = torch.ops.aten._assert_tensor_metadata.default(to_2, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_4 = None
to_4: "f32[s41, 1, s2]" = torch.ops.aten.to.dtype(to_2, torch.float32); to_2 = None
matmul: "f32[s41, 48, s2]" = torch.ops.aten.matmul.default(to_3, to_4); to_3 = to_4 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:143 in forward, code: ).transpose(1, 2)
transpose: "f32[s41, s2, 48]" = torch.ops.aten.transpose.int(matmul, 1, 2); matmul = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:144 in forward, code: emb = torch.cat((freqs, freqs), dim=-1)
cat: "f32[s41, s2, 96]" = torch.ops.aten.cat.default([transpose, transpose], -1); transpose = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:145 in forward, code: cos = emb.cos()
cos: "f32[s41, s2, 96]" = torch.ops.aten.cos.default(cat)
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:146 in forward, code: sin = emb.sin()
sin: "f32[s41, s2, 96]" = torch.ops.aten.sin.default(cat); cat = None
return (cos, sin)
Graph signature:
# inputs
p_model_embed_tokens_weight: PARAMETER target='model.embed_tokens.weight'
p_model_layers_0_self_attn_q_proj_weight: PARAMETER target='model.layers.0.self_attn.q_proj.weight'
p_model_layers_0_self_attn_k_proj_weight: PARAMETER target='model.layers.0.self_attn.k_proj.weight'
p_model_layers_0_self_attn_v_proj_weight: PARAMETER target='model.layers.0.self_attn.v_proj.weight'
p_model_layers_0_self_attn_o_proj_weight: PARAMETER target='model.layers.0.self_attn.o_proj.weight'
p_model_layers_0_mlp_gate_proj_weight: PARAMETER target='model.layers.0.mlp.gate_proj.weight'
p_model_layers_0_mlp_up_proj_weight: PARAMETER target='model.layers.0.mlp.up_proj.weight'
p_model_layers_0_mlp_down_proj_weight: PARAMETER target='model.layers.0.mlp.down_proj.weight'
p_model_layers_0_input_layernorm_weight: PARAMETER target='model.layers.0.input_layernorm.weight'
p_model_layers_0_post_attention_layernorm_weight: PARAMETER target='model.layers.0.post_attention_layernorm.weight'
p_model_norm_weight: PARAMETER target='model.norm.weight'
p_lm_head_weight: PARAMETER target='lm_head.weight'
b_model_rotary_emb_inv_freq: BUFFER target='model.rotary_emb.inv_freq' persistent=False
input_ids: USER_INPUT
attention_mask: USER_INPUT
position_ids: USER_INPUT
past_key_values_key_cache_0: USER_INPUT
past_key_values_value_cache_0: USER_INPUT
# outputs
linear_7: USER_OUTPUT
cat_3: USER_OUTPUT
cat_4: USER_OUTPUT
Range constraints: {s41: VR[1, 1024], s2: VR[2, 4096], s2 + s67: VR[4, 8192], s67: VR[1, 4096]}
If you have any error, then look at example Export Tiny-LLM with patches.
doc.plot_legend("Tiny-LLM\nforward inputs\nbehind generate", "torch.export.export", "tomato")
Total running time of the script: (0 minutes 4.838 seconds)
Related examples
