Fix Falcon model, inference test in CI (flexflow#1138)

* . * gpu torch in docker * fix * add falcon to ci * re-enabled opt tests, linting
xinhaoc · Sep 17, 2023 · c7cc6b4 · c7cc6b4
1 parent 8f04bea
commit c7cc6b4
Show file tree

Hide file tree

Showing 9 changed files with 67 additions and 53 deletions.
diff --git a/docker/flexflow-environment/Dockerfile b/docker/flexflow-environment/Dockerfile
@@ -77,7 +77,7 @@ ENV CUDA_DIR /usr/local/cuda
 # Install python packages and other dependencies
 RUN conda install -c conda-forge cmake make pillow cmake-build-extension pybind11 numpy pandas keras-preprocessing
 # Install CPU-only Pytorch and related dependencies
-RUN conda install pytorch torchvision torchaudio cpuonly -c pytorch
+RUN conda install pytorch torchvision torchaudio -c pytorch
 RUN conda install -c conda-forge onnx transformers>=4.31.0 sentencepiece einops
 RUN pip3 install tensorflow notebook
 

diff --git a/inference/python/spec_infer.py b/inference/python/spec_infer.py
@@ -67,7 +67,7 @@ def get_configs():
             "ssms": [
                 {
                     # required ssm parameter
-                    "ssm_model": "JackFram/llama-160m",
+                    "ssm_model": "JackFram/llama-160m-base",
                     # optional ssm parameters
                     "cache_path": "",
                     "refresh_cache": False,

diff --git a/src/ops/spec_inc_multihead_self_attention.cu b/src/ops/spec_inc_multihead_self_attention.cu
@@ -350,18 +350,18 @@ void compute_attention_kernel(SpecIncMultiHeadSelfAttentionMeta const *m,
       }
       // add alibi position bias to qk production
       // add alibi position bias to qk production
-    if (*m->position_bias) {
-      size_t parallelism = m->num_q_heads * total_tokens * num_new_tokens;
-      apply_position_bias_qkprd<<<GET_BLOCKS(parallelism),
-                                  min((size_t)CUDA_NUM_THREADS, parallelism),
-                                  0,
-                                  stream>>>(C,
-                                            num_new_tokens,
-                                            total_tokens,
-                                            m->num_q_heads,
-                                            m->global_num_q_heads,
-                                            shard_id);
-    }
+      if (*m->position_bias) {
+        size_t parallelism = m->num_q_heads * total_tokens * num_new_tokens;
+        apply_position_bias_qkprd<<<GET_BLOCKS(parallelism),
+                                    min((size_t)CUDA_NUM_THREADS, parallelism),
+                                    0,
+                                    stream>>>(C,
+                                              num_new_tokens,
+                                              total_tokens,
+                                              m->num_q_heads,
+                                              m->global_num_q_heads,
+                                              shard_id);
+      }
       // Fill all elements above diagonal in qk prods with -inf to force
       // causal attention.
       assert(num_new_tokens <= total_tokens);

diff --git a/src/runtime/request_manager.cc b/src/runtime/request_manager.cc
@@ -202,7 +202,7 @@ RequestManager::RequestGuid
   request.status = Request::PENDING;
   request.guid = next_available_guid++;
   request.max_sequence_length = max_sequence_length;
-  if (bos_token_id >= 0) {
+  if (bos_token_id >= 0 && model_type != ModelType::FALCON) {
     request.tokens.push_back(bos_token_id);
   }
   std::vector<int32_t> tokens = this->tokenizer_->Encode(prompt);

diff --git a/tests/inference/cpp_inference_tests.sh b/tests/inference/cpp_inference_tests.sh
@@ -10,9 +10,9 @@ cd "${BASH_SOURCE[0]%/*}"
 ###############################################################################################
 
 # LLAMA
-../../build/inference/spec_infer/spec_infer -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion --use-full-precision -llm-model decapoda-research/llama-7b-hf -ssm-model JackFram/llama-160m -prompt ../../inference/prompt/test.json -output-file ../../inference/output/spec_inference_llama.txt -pipeline-parallelism-degree 4
+../../build/inference/spec_infer/spec_infer -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion --use-full-precision -llm-model decapoda-research/llama-7b-hf -ssm-model JackFram/llama-160m-base -prompt ../../inference/prompt/test.json -output-file ../../inference/output/spec_inference_llama.txt -pipeline-parallelism-degree 4
 # LLAMA (half precision)
-../../build/inference/spec_infer/spec_infer -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion -llm-model decapoda-research/llama-7b-hf -ssm-model JackFram/llama-160m -prompt ../../inference/prompt/test.json -output-file ../../inference/output/spec_inference_llama_half.txt -pipeline-parallelism-degree 4
+../../build/inference/spec_infer/spec_infer -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion -llm-model decapoda-research/llama-7b-hf -ssm-model JackFram/llama-160m-base -prompt ../../inference/prompt/test.json -output-file ../../inference/output/spec_inference_llama_half.txt -pipeline-parallelism-degree 4
 
 # OPT
 ../../build/inference/spec_infer/spec_infer -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion --use-full-precision -llm-model facebook/opt-6.7b -ssm-model facebook/opt-125m -prompt ../../inference/prompt/test.json -output-file ../../inference/output/spec_inference_opt.txt -pipeline-parallelism-degree 4
@@ -22,9 +22,9 @@ cd "${BASH_SOURCE[0]%/*}"
 # Tensor parallelism tests
 if [ "$TENSOR_PARALLELISM_TESTS" = "ON" ]; then
     # LLAMA
-    ../../build/inference/spec_infer/spec_infer -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion --use-full-precision -llm-model decapoda-research/llama-7b-hf -ssm-model JackFram/llama-160m -prompt ../../inference/prompt/test.json -output-file ../../inference/output/spec_inference_llama_tp.txt -pipeline-parallelism-degree 2 -tensor-parallelism-degree 2
+    ../../build/inference/spec_infer/spec_infer -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion --use-full-precision -llm-model decapoda-research/llama-7b-hf -ssm-model JackFram/llama-160m-base -prompt ../../inference/prompt/test.json -output-file ../../inference/output/spec_inference_llama_tp.txt -pipeline-parallelism-degree 2 -tensor-parallelism-degree 2
     # LLAMA (half precision)
-    ../../build/inference/spec_infer/spec_infer -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion -llm-model decapoda-research/llama-7b-hf -ssm-model JackFram/llama-160m -prompt ../../inference/prompt/test.json -output-file ../../inference/output/spec_inference_llama_half_tp.txt -pipeline-parallelism-degree 2 -tensor-parallelism-degree 2
+    ../../build/inference/spec_infer/spec_infer -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion -llm-model decapoda-research/llama-7b-hf -ssm-model JackFram/llama-160m-base -prompt ../../inference/prompt/test.json -output-file ../../inference/output/spec_inference_llama_half_tp.txt -pipeline-parallelism-degree 2 -tensor-parallelism-degree 2
 
     # OPT
     ../../build/inference/spec_infer/spec_infer -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion --use-full-precision -llm-model facebook/opt-6.7b -ssm-model facebook/opt-125m -prompt ../../inference/prompt/test.json -output-file ../../inference/output/spec_inference_opt_tp.txt -pipeline-parallelism-degree 2 -tensor-parallelism-degree 2
@@ -37,9 +37,9 @@ fi
 ###############################################################################################
 
 # LLAMA (small model)
-../../build/inference/incr_decoding/incr_decoding -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion --use-full-precision -llm-model JackFram/llama-160m -prompt ../../inference/prompt/test.json -output-file ../../inference/output/incr_decoding_llama_160M.txt -pipeline-parallelism-degree 4
+../../build/inference/incr_decoding/incr_decoding -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion --use-full-precision -llm-model JackFram/llama-160m-base -prompt ../../inference/prompt/test.json -output-file ../../inference/output/incr_decoding_llama_160M.txt -pipeline-parallelism-degree 4
 # LLAMA (small model, half precision)
-../../build/inference/incr_decoding/incr_decoding -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion -llm-model JackFram/llama-160m -prompt ../../inference/prompt/test.json -output-file ../../inference/output/incr_decoding_llama_160M_half.txt -pipeline-parallelism-degree 4
+../../build/inference/incr_decoding/incr_decoding -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion -llm-model JackFram/llama-160m-base -prompt ../../inference/prompt/test.json -output-file ../../inference/output/incr_decoding_llama_160M_half.txt -pipeline-parallelism-degree 4
 
 # LLAMA (big model)
 ../../build/inference/incr_decoding/incr_decoding -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion --use-full-precision -llm-model decapoda-research/llama-7b-hf -prompt ../../inference/prompt/test.json -output-file ../../inference/output/incr_decoding_llama_7B.txt -pipeline-parallelism-degree 4
@@ -69,11 +69,11 @@ fi
 # Tensor parallelism tests
 if [ "$TENSOR_PARALLELISM_TESTS" = "ON" ]; then
     # LLAMA (small model)
-    ../../build/inference/incr_decoding/incr_decoding -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion --use-full-precision -llm-model JackFram/llama-160m -prompt ../../inference/prompt/test.json -output-file ../../inference/output/incr_decoding_llama_160M_tp.txt -pipeline-parallelism-degree 2 -tensor-parallelism-degree 2
-    ../../build/inference/incr_decoding/incr_decoding -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion --use-full-precision -llm-model JackFram/llama-160m -prompt ../../inference/prompt/test.json -output-file ../../inference/output/incr_decoding_llama_160M_tp4.txt -pipeline-parallelism-degree 1 -tensor-parallelism-degree 4
+    ../../build/inference/incr_decoding/incr_decoding -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion --use-full-precision -llm-model JackFram/llama-160m-base -prompt ../../inference/prompt/test.json -output-file ../../inference/output/incr_decoding_llama_160M_tp.txt -pipeline-parallelism-degree 2 -tensor-parallelism-degree 2
+    ../../build/inference/incr_decoding/incr_decoding -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion --use-full-precision -llm-model JackFram/llama-160m-base -prompt ../../inference/prompt/test.json -output-file ../../inference/output/incr_decoding_llama_160M_tp4.txt -pipeline-parallelism-degree 1 -tensor-parallelism-degree 4
     # LLAMA (small model, half precision)
-    ../../build/inference/incr_decoding/incr_decoding -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion -llm-model JackFram/llama-160m -prompt ../../inference/prompt/test.json -output-file ../../inference/output/incr_decoding_llama_160M_half_tp.txt -pipeline-parallelism-degree 2 -tensor-parallelism-degree 2
-    ../../build/inference/incr_decoding/incr_decoding -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion -llm-model JackFram/llama-160m -prompt ../../inference/prompt/test.json -output-file ../../inference/output/incr_decoding_llama_160M_half_tp4.txt -pipeline-parallelism-degree 1 -tensor-parallelism-degree 4
+    ../../build/inference/incr_decoding/incr_decoding -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion -llm-model JackFram/llama-160m-base -prompt ../../inference/prompt/test.json -output-file ../../inference/output/incr_decoding_llama_160M_half_tp.txt -pipeline-parallelism-degree 2 -tensor-parallelism-degree 2
+    ../../build/inference/incr_decoding/incr_decoding -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion -llm-model JackFram/llama-160m-base -prompt ../../inference/prompt/test.json -output-file ../../inference/output/incr_decoding_llama_160M_half_tp4.txt -pipeline-parallelism-degree 1 -tensor-parallelism-degree 4
 
     # LLAMA (big model)
     ../../build/inference/incr_decoding/incr_decoding -ll:gpu 4 -ll:fsize 14000 -ll:zsize 30000 --fusion --use-full-precision -llm-model decapoda-research/llama-7b-hf -prompt ../../inference/prompt/test.json -output-file ../../inference/output/incr_decoding_llama_7B_tp.txt -pipeline-parallelism-degree 2 -tensor-parallelism-degree 2
@@ -216,28 +216,32 @@ fi
 ######################### Alignment tests with HuggingFace ####################################
 
 # LLAMA (small model, full precision)
-python3 ./huggingface_inference.py --model-name "JackFram/llama-160m" --tokenizer-model-name "JackFram/llama-160m" --use-full-precision --prompt-file "../../inference/prompt/test.json" --output-file "../../inference/output/huggingface_llama_160M.txt" --gpu
+python3 ./huggingface_inference.py --model-name "JackFram/llama-160m-base" --use-full-precision --prompt-file "../../inference/prompt/test.json" --output-file "../../inference/output/huggingface_llama_160M.txt" --gpu
 
 # LLAMA (small model, half precision)
-python3 ./huggingface_inference.py --model-name "JackFram/llama-160m" --tokenizer-model-name "JackFram/llama-160m" --prompt-file "../../inference/prompt/test.json" --output-file "../../inference/output/huggingface_llama_160M_half.txt" --gpu
+python3 ./huggingface_inference.py --model-name "JackFram/llama-160m-base" --prompt-file "../../inference/prompt/test.json" --output-file "../../inference/output/huggingface_llama_160M_half.txt" --gpu
 
 # LLAMA (big model, full precision)
-python3 ./huggingface_inference.py --model-name "decapoda-research/llama-7b-hf" --tokenizer-model-name "JackFram/llama-160m" --use-full-precision --prompt-file "../../inference/prompt/test.json" --output-file "../../inference/output/huggingface_llama_7B.txt"
+python3 ./huggingface_inference.py --model-name "decapoda-research/llama-7b-hf" --use-full-precision --prompt-file "../../inference/prompt/test.json" --output-file "../../inference/output/huggingface_llama_7B.txt"
 
 # LLAMA (big model, half precision)
-python3 ./huggingface_inference.py --model-name "decapoda-research/llama-7b-hf" --tokenizer-model-name "JackFram/llama-160m" --prompt-file "../../inference/prompt/test.json" --output-file "../../inference/output/huggingface_llama_7B_half.txt" --gpu
+python3 ./huggingface_inference.py --model-name "decapoda-research/llama-7b-hf" --prompt-file "../../inference/prompt/test.json" --output-file "../../inference/output/huggingface_llama_7B_half.txt" --gpu
 
 # OPT (small model, full precision)
-python3 ./huggingface_inference.py --model-name "facebook/opt-125m" --tokenizer-model-name "facebook/opt-125m" --use-full-precision --prompt-file "../../inference/prompt/test.json" --output-file "../../inference/output/huggingface_opt_125M.txt" --gpu --max-length 128
+python3 ./huggingface_inference.py --model-name "facebook/opt-125m" --use-full-precision --prompt-file "../../inference/prompt/test.json" --output-file "../../inference/output/huggingface_opt_125M.txt" --gpu --max-length 128
 
 # OPT (small model, half precision)
-python3 ./huggingface_inference.py --model-name "facebook/opt-125m" --tokenizer-model-name "facebook/opt-125m" --prompt-file "../../inference/prompt/test.json" --output-file "../../inference/output/huggingface_opt_125M_half.txt" --gpu --max-length 128
+python3 ./huggingface_inference.py --model-name "facebook/opt-125m" --prompt-file "../../inference/prompt/test.json" --output-file "../../inference/output/huggingface_opt_125M_half.txt" --gpu --max-length 128
 
 # OPT (big model, full precision)
-#python3 ./huggingface_inference.py --model-name "facebook/opt-6.7b" --tokenizer-model-name "facebook/opt-6.7b" --use-full-precision --prompt-file "../../inference/prompt/test.json" --output-file "../../inference/output/huggingface_opt_6B.txt" --max-length 127
+python3 ./huggingface_inference.py --model-name "facebook/opt-6.7b" --use-full-precision --prompt-file "../../inference/prompt/test.json" --output-file "../../inference/output/huggingface_opt_6B.txt" --max-length 128
 
 # OPT (big model, half precision)
-#python3 ./huggingface_inference.py --model-name "facebook/opt-6.7b" --tokenizer-model-name "facebook/opt-6.7b" --prompt-file "../../inference/prompt/test.json" --output-file "../../inference/output/huggingface_opt_6B_half.txt" --gpu --max-length 127
+# python3 ./huggingface_inference.py --model-name "facebook/opt-6.7b" --prompt-file "../../inference/prompt/test.json" --output-file "../../inference/output/huggingface_opt_6B_half.txt" --gpu --max-length 128
+
+# Falcon (full precision)
+python3 ./huggingface_inference.py --model-name "tiiuae/falcon-7b" --use-full-precision --prompt-file "../../inference/prompt/test.json" --output-file "../../inference/output/huggingface_falcon_7B.txt" --max-length 128
+
 
 diff <(tail -n +2 "../../inference/output/huggingface_llama_160M.txt") <(tail -n +5 "../../inference/output/incr_decoding_llama_160M.txt")
 diff <(tail -n +2 "../../inference/output/huggingface_llama_160M_half.txt" | tr -s '[:space:]' '\n' | head -n 20) <(tail -n +5 "../../inference/output/incr_decoding_llama_160M_half.txt" | tr -s '[:space:]' '\n' | head -n 20)
@@ -246,5 +250,7 @@ diff <(tail -n +2 "../../inference/output/huggingface_llama_7B_half.txt" | tr -s
 
 diff <(tail -n +2 "../../inference/output/huggingface_opt_125M.txt") <(tail -n +5 "../../inference/output/incr_decoding_opt_125M.txt")
 diff <(tail -n +2 "../../inference/output/huggingface_opt_125M_half.txt" | tr -s '[:space:]' '\n' | head -n 20) <(tail -n +5 "../../inference/output/incr_decoding_opt_125M_half.txt" | tr -s '[:space:]' '\n' | head -n 20)
-#diff <(tail -n +2 "../../inference/output/huggingface_opt_6B.txt") <(tail -n +5 "../../inference/output/incr_decoding_opt_6B.txt")
-#diff <(tail -n +2 "../../inference/output/huggingface_opt_6B_half.txt") <(tail -n +5 "../../inference/output/incr_decoding_opt_6B_half.txt")
+diff <(tail -n +2 "../../inference/output/huggingface_opt_6B.txt") <(tail -n +5 "../../inference/output/incr_decoding_opt_6B.txt")
+# diff <(tail -n +2 "../../inference/output/huggingface_opt_6B_half.txt") <(tail -n +5 "../../inference/output/incr_decoding_opt_6B_half.txt")
+diff <(tail -n +2 "../../inference/output/huggingface_falcon_7B.txt") <(tail -n +5 "../../inference/output/incr_decoding_falcon_7B.txt")
+
diff --git a/tests/inference/huggingface_inference.py b/tests/inference/huggingface_inference.py
@@ -1,7 +1,7 @@
 import argparse
 import json
 import os
-from transformers import AutoModelForCausalLM, AutoTokenizer, LlamaTokenizer
+from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig, LlamaTokenizer
 
 def main():
     # Change working dir to folder storing this script
@@ -12,7 +12,6 @@ def main():
     # Parse command line arguments
     parser = argparse.ArgumentParser()
     parser.add_argument("--model-name", type=str, required=True)
-    parser.add_argument("--tokenizer-model-name", type=str, required=True)
     parser.add_argument("--max-length", type=int, default=128)
     parser.add_argument("--prompt-file", type=str, required=True)
     parser.add_argument("--output-file", type=str, required=True)
@@ -46,15 +45,20 @@ def main():
 
     # Run huggingface model
     device = "cuda" if args.gpu else "cpu"
+    # Get Model
     model = AutoModelForCausalLM.from_pretrained(args.model_name).to(device)
-    if args.tokenizer_model_name == "JackFram/llama-160m":
-        tokenizer = LlamaTokenizer.from_pretrained("JackFram/llama-160m", use_fast=True)
+    # Get Tokenizer
+    hf_config = AutoConfig.from_pretrained(args.model_name, trust_remote_code=True)
+    hf_arch = getattr(hf_config, "architectures")[0]
+    if hf_arch == "LLaMAForCausalLM" or hf_arch == "LlamaForCausalLM":
+        tokenizer = LlamaTokenizer.from_pretrained(args.model_name, use_fast=True)
     else:
-        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_model_name)
+        tokenizer = AutoTokenizer.from_pretrained(args.model_name)
+    # Generate output
     with open(args.output_file, "w") as f:
         for i, prompt in enumerate(prompt_list):
             batch = tokenizer(
-                prompt_list, return_tensors="pt", add_special_tokens=True
+                prompt, return_tensors="pt", add_special_tokens=True
             ).to(device)
             generated = model.generate(batch["input_ids"], max_length=args.max_length)
             out = tokenizer.decode(generated[0])