update readme and add artifacts

README.md

@@ -65,24 +65,28 @@ This model was obtained by quantizing the weights and activations of DeepSeek R1
 
 ### Deploy with TensorRT-LLM
 
-To deploy the quantized checkpoint with [TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM) LLM API, follow the sample codes below:
+To deploy the quantized FP4 checkpoint with [TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM) LLM API, follow the sample codes below (you need 8xB200 GPU and TensorRT-LLM 0.18 or install by building from source with the latest main branch):
 
 * LLM API sample usage:
 ```
-from tensorrt_llm import LLM, SamplingParams
+from tensorrt_llm import SamplingParams
+from tensorrt_llm._torch import LLM
+from tensorrt_llm._torch.pyexecutor.config import PyTorchConfig
 
 
 def main():
 
+    pytorch_config = PyTorchConfig()
+
     prompts = [
         "Hello, my name is",
         "The president of the United States is",
         "The capital of France is",
         "The future of AI is",
     ]
-    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
+    sampling_params = SamplingParams(max_tokens=32)
 
-    llm = LLM(model="nvidia/DeepSeek-R1-FP4")
+    llm = LLM(model="nvidia/DeepSeek-R1-FP4", tensor_parallel_size=8, pytorch_backend_config=pytorch_config, enable_attention_dp=True)
 
     outputs = llm.generate(prompts, sampling_params)
 
@@ -111,7 +115,7 @@ tar -xf data/mmlu.tar -C data && mv data/data data/mmlu
 2) Measure MMLU:
 
 ```sh
-python examples/mmlu_llmapi.py --data_dir data/mmlu --hf_model_dir nvidia/DeepSeek-R1-FP4 --backend=pytorch
+python examples/mmlu_llmapi.py --data_dir data/mmlu --hf_model_dir nvidia/DeepSeek-R1-FP4 --tp_size 8 --backend=pytorch
 ```
 
 * Throughputs evaluation:

config.json

@@ -0,0 +1,70 @@
+{
+  "architectures": [
+    "DeepseekV3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_deepseek.DeepseekV3Config",
+    "AutoModel": "modeling_deepseek.DeepseekV3Model",
+    "AutoModelForCausalLM": "modeling_deepseek.DeepseekV3ForCausalLM"
+  },
+  "aux_loss_alpha": 0.001,
+  "bos_token_id": 0,
+  "eos_token_id": 1,
+  "ep_size": 1,
+  "first_k_dense_replace": 3,
+  "hidden_act": "silu",
+  "hidden_size": 7168,
+  "initializer_range": 0.02,
+  "intermediate_size": 18432,
+  "kv_lora_rank": 512,
+  "max_position_embeddings": 163840,
+  "model_type": "deepseek_v3",
+  "moe_intermediate_size": 2048,
+  "moe_layer_freq": 1,
+  "n_group": 8,
+  "n_routed_experts": 256,
+  "n_shared_experts": 1,
+  "norm_topk_prob": true,
+  "num_attention_heads": 128,
+  "num_experts_per_tok": 8,
+  "num_hidden_layers": 61,
+  "num_key_value_heads": 128,
+  "num_nextn_predict_layers": 1,
+  "pretraining_tp": 1,
+  "q_lora_rank": 1536,
+  "qk_nope_head_dim": 128,
+  "qk_rope_head_dim": 64,
+  "quantization_config": {
+    "activation_scheme": "dynamic",
+    "fmt": "e4m3",
+    "quant_method": "fp8",
+    "weight_block_size": [
+      128,
+      128
+    ]
+  },
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": {
+    "beta_fast": 32,
+    "beta_slow": 1,
+    "factor": 40,
+    "mscale": 1.0,
+    "mscale_all_dim": 1.0,
+    "original_max_position_embeddings": 4096,
+    "type": "yarn"
+  },
+  "rope_theta": 10000,
+  "routed_scaling_factor": 2.5,
+  "scoring_func": "sigmoid",
+  "seq_aux": true,
+  "tie_word_embeddings": false,
+  "topk_group": 4,
+  "topk_method": "noaux_tc",
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.46.3",
+  "use_cache": true,
+  "v_head_dim": 128,
+  "vocab_size": 129280
+}

generation_config.json

@@ -0,0 +1,9 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 0,
+  "eos_token_id": 1,
+  "do_sample": true,
+  "temperature": 0.6,
+  "top_p": 0.95,
+  "transformers_version": "4.39.3"
+}

hf_quant_config.json

@@ -0,0 +1,258 @@
+{
+    "producer": {
+        "name": "modelopt",
+        "version": "0.23.0"
+    },
+    "quantization": {
+        "quant_algo": "NVFP4",
+        "kv_cache_quant_algo": null,
+        "group_size": 16,
+        "exclude_modules": [
+            "model.layers.30.self_attn*",
+            "model.layers.33.input_layernorm",
+            "model.layers.39.input_layernorm",
+            "model.layers.26.mlp.gate",
+            "model.layers.36.mlp.gate",
+            "model.layers.16.input_layernorm",
+            "model.layers.45.post_attention_layernorm",
+            "model.layers.52.input_layernorm",
+            "model.layers.38.input_layernorm",
+            "model.layers.28.post_attention_layernorm",
+            "model.layers.59.self_attn*",
+            "model.layers.25.self_attn*",
+            "model.layers.34.input_layernorm",
+            "model.layers.44.mlp.gate",
+            "model.layers.8.mlp.gate",
+            "model.layers.59.input_layernorm",
+            "model.layers.2.self_attn*",
+            "model.layers.40.input_layernorm",
+            "model.layers.11.post_attention_layernorm",
+            "model.layers.40.mlp.gate",
+            "model.layers.5.self_attn*",
+            "model.layers.29.post_attention_layernorm",
+            "model.layers.54.self_attn*",
+            "model.layers.3.input_layernorm",
+            "model.layers.43.self_attn*",
+            "model.layers.51.post_attention_layernorm",
+            "model.layers.50.post_attention_layernorm",
+            "model.norm",
+            "model.layers.57.input_layernorm",
+            "model.layers.22.self_attn*",
+            "model.layers.27.self_attn*",
+            "model.layers.21.self_attn*",
+            "model.layers.9.mlp.gate",
+            "model.layers.60.self_attn*",
+            "model.layers.24.post_attention_layernorm",
+            "model.layers.27.input_layernorm",
+            "model.layers.1.input_layernorm",
+            "model.layers.3.post_attention_layernorm",
+            "model.layers.57.mlp.gate",
+            "model.layers.18.input_layernorm",
+            "model.layers.58.mlp.gate",
+            "model.layers.33.mlp.gate",
+            "model.layers.52.post_attention_layernorm",
+            "model.layers.10.mlp.gate",
+            "model.layers.44.self_attn*",
+            "model.layers.29.input_layernorm",
+            "model.layers.15.self_attn*",
+            "model.layers.21.mlp.gate",
+            "model.layers.48.post_attention_layernorm",
+            "model.layers.12.mlp.gate",
+            "model.layers.30.input_layernorm",
+            "model.layers.34.post_attention_layernorm",
+            "model.layers.41.self_attn*",
+            "model.layers.18.mlp.gate",
+            "model.layers.24.mlp.gate",
+            "model.layers.42.mlp.gate",
+            "model.layers.0.input_layernorm",
+            "model.layers.23.self_attn*",
+            "model.layers.20.mlp.gate",
+            "model.layers.6.mlp.gate",
+            "model.layers.52.self_attn*",
+            "model.layers.30.post_attention_layernorm",
+            "model.layers.35.post_attention_layernorm",
+            "model.layers.11.self_attn*",
+            "model.layers.23.post_attention_layernorm",
+            "model.layers.51.mlp.gate",
+            "model.layers.4.mlp.gate",
+            "model.layers.22.mlp.gate",
+            "model.layers.41.post_attention_layernorm",
+            "model.layers.6.input_layernorm",
+            "model.layers.53.mlp.gate",
+            "model.layers.46.post_attention_layernorm",
+            "model.layers.12.post_attention_layernorm",
+            "model.layers.13.input_layernorm",
+            "model.layers.4.self_attn*",
+            "model.layers.29.self_attn*",
+            "model.layers.45.self_attn*",
+            "model.layers.9.self_attn*",
+            "model.layers.56.mlp.gate",
+            "model.layers.31.post_attention_layernorm",
+            "model.layers.47.mlp.gate",
+            "model.layers.49.post_attention_layernorm",
+            "model.layers.7.input_layernorm",
+            "model.layers.20.post_attention_layernorm",
+            "model.layers.14.post_attention_layernorm",
+            "model.layers.11.input_layernorm",
+            "model.layers.12.self_attn*",
+            "model.layers.42.input_layernorm",
+            "model.layers.26.self_attn*",
+            "model.layers.43.post_attention_layernorm",
+            "model.layers.23.input_layernorm",
+            "model.layers.16.mlp.gate",
+            "model.layers.31.mlp.gate",
+            "model.layers.50.mlp.gate",
+            "model.layers.46.input_layernorm",
+            "model.layers.40.post_attention_layernorm",
+            "model.layers.1.post_attention_layernorm",
+            "model.layers.53.input_layernorm",
+            "model.layers.39.self_attn*",
+            "model.layers.27.post_attention_layernorm",
+            "model.layers.16.self_attn*",
+            "model.layers.33.self_attn*",
+            "model.layers.8.input_layernorm",
+            "model.layers.59.post_attention_layernorm",
+            "model.layers.37.input_layernorm",
+            "model.layers.22.post_attention_layernorm",
+            "model.layers.7.mlp.gate",
+            "model.layers.0.self_attn*",
+            "model.layers.37.self_attn*",
+            "model.layers.3.mlp.gate",
+            "model.layers.55.self_attn*",
+            "model.layers.2.post_attention_layernorm",
+            "model.layers.19.mlp.gate",
+            "model.layers.13.mlp.gate",
+            "model.layers.7.self_attn*",
+            "model.layers.47.input_layernorm",
+            "model.layers.32.mlp.gate",
+            "model.layers.10.input_layernorm",
+            "model.layers.50.input_layernorm",
+            "model.layers.51.input_layernorm",
+            "model.layers.55.post_attention_layernorm",
+            "model.layers.4.post_attention_layernorm",
+            "model.layers.20.input_layernorm",
+            "model.layers.45.input_layernorm",
+            "model.layers.49.self_attn*",
+            "model.layers.22.input_layernorm",
+            "model.layers.60.input_layernorm",
+            "model.layers.28.mlp.gate",
+            "model.layers.57.post_attention_layernorm",
+            "model.layers.51.self_attn*",
+            "model.layers.56.input_layernorm",
+            "model.layers.18.self_attn*",
+            "model.layers.11.mlp.gate",
+            "model.layers.17.input_layernorm",
+            "model.layers.14.self_attn*",
+            "model.layers.56.self_attn*",
+            "model.layers.15.post_attention_layernorm",
+            "model.layers.19.self_attn*",
+            "lm_head",
+            "model.layers.40.self_attn*",
+            "model.layers.41.input_layernorm",
+            "model.layers.44.input_layernorm",
+            "model.layers.25.mlp.gate",
+            "model.layers.12.input_layernorm",
+            "model.layers.53.post_attention_layernorm",
+            "model.layers.2.input_layernorm",
+            "model.layers.19.post_attention_layernorm",
+            "model.layers.48.input_layernorm",
+            "model.layers.31.self_attn*",
+            "model.layers.14.mlp.gate",
+            "model.layers.30.mlp.gate",
+            "model.layers.60.post_attention_layernorm",
+            "model.layers.41.mlp.gate",
+            "model.layers.1.self_attn*",
+            "model.layers.52.mlp.gate",
+            "model.layers.29.mlp.gate",
+            "model.layers.14.input_layernorm",
+            "model.layers.5.post_attention_layernorm",
+            "model.layers.23.mlp.gate",
+            "model.layers.42.post_attention_layernorm",
+            "model.layers.35.input_layernorm",
+            "model.layers.17.self_attn*",
+            "model.layers.28.self_attn*",
+            "model.layers.58.self_attn*",
+            "model.layers.13.post_attention_layernorm",
+            "model.layers.32.post_attention_layernorm",
+            "model.layers.10.self_attn*",
+            "model.layers.33.post_attention_layernorm",
+            "model.layers.38.mlp.gate",
+            "model.layers.5.input_layernorm",
+            "model.layers.26.post_attention_layernorm",
+            "model.layers.15.mlp.gate",
+            "model.layers.25.input_layernorm",
+            "model.layers.9.post_attention_layernorm",
+            "model.layers.43.input_layernorm",
+            "model.layers.47.self_attn*",
+            "model.layers.32.self_attn*",
+            "model.layers.61*",
+            "model.layers.35.self_attn*",
+            "model.layers.24.self_attn*",
+            "model.layers.46.self_attn*",
+            "model.layers.13.self_attn*",
+            "model.layers.53.self_attn*",
+            "model.layers.43.mlp.gate",
+            "model.layers.55.mlp.gate",
+            "model.layers.54.post_attention_layernorm",
+            "model.layers.18.post_attention_layernorm",
+            "model.layers.31.input_layernorm",
+            "model.layers.6.self_attn*",
+            "model.layers.17.post_attention_layernorm",
+            "model.layers.24.input_layernorm",
+            "model.layers.20.self_attn*",
+            "model.layers.36.post_attention_layernorm",
+            "model.layers.32.input_layernorm",
+            "model.layers.28.input_layernorm",
+            "model.layers.26.input_layernorm",
+            "model.layers.36.self_attn*",
+            "model.layers.0.post_attention_layernorm",
+            "model.layers.39.post_attention_layernorm",
+            "model.layers.56.post_attention_layernorm",
+            "model.layers.39.mlp.gate",
+            "model.layers.9.input_layernorm",
+            "model.layers.54.mlp.gate",
+            "model.layers.5.mlp.gate",
+            "model.layers.16.post_attention_layernorm",
+            "model.layers.55.input_layernorm",
+            "model.layers.46.mlp.gate",
+            "model.layers.57.self_attn*",
+            "model.layers.10.post_attention_layernorm",
+            "model.layers.48.self_attn*",
+            "model.layers.21.input_layernorm",
+            "model.layers.44.post_attention_layernorm",
+            "model.layers.17.mlp.gate",
+            "model.layers.37.post_attention_layernorm",
+            "model.layers.49.input_layernorm",
+            "model.layers.49.mlp.gate",
+            "model.layers.15.input_layernorm",
+            "model.layers.45.mlp.gate",
+            "model.layers.38.self_attn*",
+            "model.layers.47.post_attention_layernorm",
+            "model.layers.37.mlp.gate",
+            "model.layers.25.post_attention_layernorm",
+            "model.embed_tokens",
+            "model.layers.36.input_layernorm",
+            "model.layers.38.post_attention_layernorm",
+            "model.layers.35.mlp.gate",
+            "model.layers.59.mlp.gate",
+            "model.layers.50.self_attn*",
+            "model.layers.54.input_layernorm",
+            "model.layers.58.input_layernorm",
+            "model.layers.21.post_attention_layernorm",
+            "model.layers.3.self_attn*",
+            "model.layers.58.post_attention_layernorm",
+            "model.layers.34.mlp.gate",
+            "model.layers.6.post_attention_layernorm",
+            "model.layers.34.self_attn*",
+            "model.layers.7.post_attention_layernorm",
+            "model.layers.42.self_attn*",
+            "model.layers.19.input_layernorm",
+            "model.layers.48.mlp.gate",
+            "model.layers.4.input_layernorm",
+            "model.layers.27.mlp.gate",
+            "model.layers.8.self_attn*",
+            "model.layers.8.post_attention_layernorm",
+            "model.layers.60.mlp.gate"
+        ]
+    }
+}

tokenizer_config.json

@@ -0,0 +1,35 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<｜begin▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "clean_up_tokenization_spaces": false,
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "<｜end▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "legacy": true,
+  "model_max_length": 16384,
+  "pad_token": {
+    "__type": "AddedToken",
+    "content": "<｜end▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sp_model_kwargs": {},
+  "unk_token": null,
+  "tokenizer_class": "LlamaTokenizerFast",
+  "chat_template": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='', is_first_sp=true) %}{%- for message in messages %}{%- if message['role'] == 'system' %}{%- if ns.is_first_sp %}{% set ns.system_prompt = ns.system_prompt + message['content'] %}{% set ns.is_first_sp = false %}{%- else %}{% set ns.system_prompt = ns.system_prompt + '\\n\\n' + message['content'] %}{%- endif %}{%- endif %}{%- endfor %}{{ bos_token }}{{ ns.system_prompt }}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and 'tool_calls' in message %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls'] %}{%- if not ns.is_first %}{%- if message['content'] is none %}{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- else %}{{'<｜Assistant｜>' + message['content'] + '<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- endif %}{%- set ns.is_first = true -%}{%- else %}{{'\\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- endif %}{%- endfor %}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- if message['role'] == 'assistant' and 'tool_calls' not in message %}{%- if ns.is_tool %}{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{% set content = message['content'] %}{% if '</think>' in content %}{% set content = content.split('</think>')[-1] %}{% endif %}{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_tool = true -%}{%- if ns.is_output_first %}{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- set ns.is_output_first = false %}{%- else %}{{'<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- endif %}{%- endfor -%}{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}{% endif %}{% if add_generation_prompt and not ns.is_tool %}{{'<｜Assistant｜><think>\\n'}}{% endif %}"
+}