examples(text-generation): update numbers for streamlined models

examples/nlp/text-generation/README.md

@@ -4,25 +4,28 @@ The model is quantized and its ability to predict the last token of some samples
 
 Models using `int8` or `float8` weights are calibrated using `128` samples of the `lambada` validation dataset.
 
-Note: since the calibration samples are shuffled, the results might be slightly different between runs.
+Model activations are streamlined: if the next operation is incompatible with quantized inputs, then the activations are disabled
+to avoid a spurious quantize/dequantize that only degrades accuracy.
+
+Note: since the calibration samples are shuffled, the results might be slightly different between runs if you change the seed.
 
 | model                            | fp16 | w int8 a fp16 | w int8 a int8 | w int8 a fp8_e5m2 | w int8 a fp8_e4m3 |
 |----------------------------------|------|---------------|---------------|-------------------|-------------------|
-| facebook/opt-125m                | 0.63 | 0.63          | 0.05          | 0.01              | **0.61**          |
-| facebook/opt-350m                | 0.67 | 0.67          | **0.63**      | 0.00              | **0.64**          |
-| facebook/opt-1.3b                | 0.76 | 0.76          | 0.51          | 0.00              | **0.71**          |
+| facebook/opt-125m                | 0.63 | 0.63          | 0.06          | 0.00              | **0.61**          |
+| facebook/opt-350m                | 0.67 | 0.67          | **0.64**      | 0.00              | **0.65**          |
+| facebook/opt-1.3b                | 0.76 | 0.76          | 0.18          | 0.00              | **0.70**          |
 | EleutherAI/pythia-160m           | 0.44 | 0.44          | 0.00          | 0.00              | 0.00              |
-| EleutherAI/pythia-410m           | 0.68 | 0.68          | 0.12          | 0.05              | 0.32              |
-| EleutherAI/pythia-1b             | 0.71 | 0.72          | 0.42          | 0.45              | **0.67**          |
-| princeton-nlp/Sheared-LLaMA-1.3B | 0.83 | 0.83          | **0.71**      | **0.75**          | **0.80**          |
-| 01-ai/Yi-6B (bfloat 16)          | 0.82 | 0.82          | 0.25          | 0.68              | **0.76**          |
-| NousResearch/Llama-2-7b-hf       | 0.89 | 0.89          | **0.77**      | 0.04              | **0.85**          |
-| HuggingFaceH4/zephyr-7b-beta     | 0.86 | 0.86          | 0.31          | **0.79**          | **0.83**          |
+| EleutherAI/pythia-410m           | 0.68 | 0.68          | 0.43          | 0.22              | 0.41              |
+| EleutherAI/pythia-1b             | 0.71 | 0.72          | **0.70**      | **0.69**          | **0.71**          |
+| princeton-nlp/Sheared-LLaMA-1.3B | 0.83 | 0.83          | **0.81**      | **0.72**          | **0.80**          |
+| 01-ai/Yi-6B (bfloat 16)          | 0.83 | 0.83          | **0.80**      | 0.00              | **0.76**          |
+| NousResearch/Llama-2-7b-hf       | 0.89 | 0.89          | **0.87**      | **0.76**          | **0.85**          |
+| HuggingFaceH4/zephyr-7b-beta     | 0.86 | 0.86          | **0.85**      | 0.00              | **0.83**          |
 
 As we can see, there is no performance degradation when quantizing only the weights to int8.
 
 When quantizing also the activations, there are wild variations between models, but we can see a pattern:
 - smaller models are generally more sensitive to 8-bit activations,
-- for models whose accuracy is degraded, using float8 activations always gives better result.
+- for models whose accuracy is degraded, using float8 activations often gives better result.
 
 Some of the models (**in bold**) are doing extremely well.